{"exhaustive":{"nbHits":false,"typo":false},"exhaustiveNbHits":false,"exhaustiveTypo":false,"hits":[{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"stephensonsco"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["deepgram"],"value":"Show HN: Kur by Deepgram (YC S16) \u2013 Deep Learning Made Easy"},"url":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["deepgram"],"value":"http://kur.deepgram.com"}},"_tags":["story","author_stephensonsco","story_13425089","show_hn"],"author":"stephensonsco","children":[13425122,13425139,13427536,13429782,13431238,13449924],"created_at":"2017-01-18T09:20:06Z","created_at_i":1484731206,"num_comments":25,"objectID":"13425089","points":103,"story_id":13425089,"title":"Show HN: Kur by Deepgram (YC S16) \u2013 Deep Learning Made Easy","updated_at":"2024-09-20T00:14:03Z","url":"http://kur.deepgram.com"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"pain_perdu"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["deepgram"],"value":"Deepgram \u2013 Find Damning Soundbites"},"url":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["deepgram"],"value":"http://blog.deepgram.com/new-tech-lets-journalists-find-damning-soundbites/"}},"_tags":["story","author_pain_perdu","story_12674572"],"author":"pain_perdu","children":[12674952,12674990,12675011,12675046,12675173,12675178,12675580,12675616,12675984,12676381,12676460,12677550],"created_at":"2016-10-10T01:56:27Z","created_at_i":1476064587,"num_comments":56,"objectID":"12674572","points":94,"story_id":12674572,"title":"Deepgram \u2013 Find Damning Soundbites","updated_at":"2024-09-19T23:50:09Z","url":"http://blog.deepgram.com/new-tech-lets-journalists-find-damning-soundbites/"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"Alyx1337"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["deepgram"],"value":"Jarvis: A Voice Virtual Assistant in Python (OpenAI, ElevenLabs, Deepgram)"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://github.com/AlexandreSajus/JARVIS"}},"_tags":["story","author_Alyx1337","story_38682095"],"author":"Alyx1337","children":[38682096,38682648,38682698,38682943,38683105,38683180,38683230,38683347,38683374,38683377,38683428,38684517,38685045,38686879],"created_at":"2023-12-18T13:27:35Z","created_at_i":1702906055,"num_comments":58,"objectID":"38682095","points":83,"story_id":38682095,"title":"Jarvis: A Voice Virtual Assistant in Python (OpenAI, ElevenLabs, Deepgram)","updated_at":"2024-09-20T15:56:24Z","url":"https://github.com/AlexandreSajus/JARVIS"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"stephensonsco"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["deepgram"],"value":"Hey HN,

I\u2019m Scott Stephenson, one of the cofounders of Deepgram (https://www.deepgram.com/). Getting information from recorded phone calls and meetings is time-intensive, costly, and imprecise. Our speech recognition API allows businesses to reliably translate high-value unstructured audio into accurate, parsable data.

Deepgram started when my cofounder Noah Shutty and I had just finished looking for dark matter (while in a particle physics lab at University of Michigan). Noah had the idea to start recording all audio from his life, 24/7. After gathering hundreds of hours of recordings, we wanted to search inside this fresh dataset, but realized there wasn\u2019t a good way to find specific moments. So, we built a tool utilizing the same AI techniques we used for finding dark matter particle events, and it ended up working pretty well. A few months later, we made a single page demo to show off \u201csearching through sound\u201d and posted to HN. Pretty soon we were in the winter batch of YC in 2016 (https://techcrunch.com/2016/09/27/launching-a-google-for-sou...).

I\u2019d say we didn\u2019t know what we were getting ourselves into. Speech is a really big problem with a huge market, but it\u2019s also a tough nut to crack. For decades, companies have been unable to get real learnings from their massive amounts of recorded audio (some companies record more than 1,000,000 minutes of call center calls every single day). They have a few reasons why they record the audio \u2014 some for compliance, some for training, and some for market research. The questions they\u2019re trying to answer are usually as simple as:

  - \u201cWhat is the topic of the call?\u201d \n  - \u201cIs this call compliant?\u201d (did I say: my company name, my name, and \u201cthis call may be recorded\u201d)\n  - \u201cAre people getting their problems solved quickly?\u201d \n  - \u201cDo my agents need training?\u201d \n  - \u201cWhat are our customers talking about? Competitors? Our latest marketing campaign?\u201d\n\n
\nIt\u2019s the most intimate view you can get on your customers, but the problem is so large and difficult to solve that companies pushed it into the corner over the past couple decades, only trying to mitigate the bleeding. Current tools only transcribe with around 50-60% accuracy on real-world, noisy, accented, industry-specific audio (don\u2019t believe the \u2018human level accuracy\u2019 hype). When companies start solving problems using speech data, they first want transcription that\u2019s accurate. After accuracy, comes scale \u2014 another big problem. Speech processing is computationally expensive and slow. Imagine trying to get into an iterative problem solving loop when you have to wait 24 hours to get your transcripts back.

So we\u2019ve set our sights on building the speech company. Competition from companies like Google, Amazon, and Nuance is real, but none of these approach speech recognition like we do. We've rebuilt the entire speech processing stack, replacing heuristics and stats based speech processing with fully end-to-end deep learning (we use CNNs and RNNs). Using GPUs, we train speech models to learn customer\u2019s unique vocabularies, accents, product names, and acoustic environments. This can be the difference between correctly capturing \u201cwasn\u2019t delivered\u201d and \u201cwas in the liver.\u201d We\u2019ve focused on speed since we think that\u2019s very important for exploration and scale. Our API returns hour-long transcripts interactively in seconds. It\u2019s a tool many businesses wish they had.

So far we\u2019ve released tools that:

  - transcribe speech with timestamps\n  - support real-time streaming\n  - have multi-channel support\n  - understand multiple languages (in beta now)\n  - allow you to deeply search for keywords and phrases\n  - transcribe to phonemes\n  - get more accurate with use\n
\nSome of those are better mousetraps of things you\u2019re familiar with and some are completely new levers to pull in your audio data. We\u2019ve built the core on English but now we\u2019re releasing the tools for all of the Americas. (aside: You can transfer learn speech and it works well!)

Accuracy will continue to improve for transcription, but I think we can do more. It's such a large problem, and we really want to make a dent in \u201csolving speech\u201d. That means asking, truly: \u201cWhat can a human do?\u201c

People can, with little context, jump into a conversation and determine:

  - What are the words? When are they said? Who said what?\n  - Is this person young/old? Male/Female? Exhausted/energetic?\n  - Where is there confusion?\n  - What language are they speaking? What\u2019s the speaker\u2019s accent?\n  - What\u2019s the topic of the conversation? Small talk or real? Is it going well?\n
\nSome of those things are being worked on now: additional language support, language and accent detection, sentiment analysis, auto-summarization, topic modeling, and more.

We\u2019d love to hear your feedback and ideas."},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["deepgram"],"value":"Launch HN: Deepgram (YC W16) \u2013 Scalable Speech API for Businesses"}},"_tags":["story","author_stephensonsco","story_18708876","launch_hn"],"author":"stephensonsco","children":[18709564,18709787,18709964,18710328,18710336,18710419,18710881,18711276,18727406],"created_at":"2018-12-18T18:06:47Z","created_at_i":1545156407,"num_comments":23,"objectID":"18708876","points":56,"story_id":18708876,"story_text":"Hey HN,

I\u2019m Scott Stephenson, one of the cofounders of Deepgram (https://www.deepgram.com/). Getting information from recorded phone calls and meetings is time-intensive, costly, and imprecise. Our speech recognition API allows businesses to reliably translate high-value unstructured audio into accurate, parsable data.

Deepgram started when my cofounder Noah Shutty and I had just finished looking for dark matter (while in a particle physics lab at University of Michigan). Noah had the idea to start recording all audio from his life, 24/7. After gathering hundreds of hours of recordings, we wanted to search inside this fresh dataset, but realized there wasn\u2019t a good way to find specific moments. So, we built a tool utilizing the same AI techniques we used for finding dark matter particle events, and it ended up working pretty well. A few months later, we made a single page demo to show off \u201csearching through sound\u201d and posted to HN. Pretty soon we were in the winter batch of YC in 2016 (https://techcrunch.com/2016/09/27/launching-a-google-for-sou...).

I\u2019d say we didn\u2019t know what we were getting ourselves into. Speech is a really big problem with a huge market, but it\u2019s also a tough nut to crack. For decades, companies have been unable to get real learnings from their massive amounts of recorded audio (some companies record more than 1,000,000 minutes of call center calls every single day). They have a few reasons why they record the audio \u2014 some for compliance, some for training, and some for market research. The questions they\u2019re trying to answer are usually as simple as:

  - \u201cWhat is the topic of the call?\u201d \n  - \u201cIs this call compliant?\u201d (did I say: my company name, my name, and \u201cthis call may be recorded\u201d)\n  - \u201cAre people getting their problems solved quickly?\u201d \n  - \u201cDo my agents need training?\u201d \n  - \u201cWhat are our customers talking about? Competitors? Our latest marketing campaign?\u201d\n\n
\nIt\u2019s the most intimate view you can get on your customers, but the problem is so large and difficult to solve that companies pushed it into the corner over the past couple decades, only trying to mitigate the bleeding. Current tools only transcribe with around 50-60% accuracy on real-world, noisy, accented, industry-specific audio (don\u2019t believe the \u2018human level accuracy\u2019 hype). When companies start solving problems using speech data, they first want transcription that\u2019s accurate. After accuracy, comes scale \u2014 another big problem. Speech processing is computationally expensive and slow. Imagine trying to get into an iterative problem solving loop when you have to wait 24 hours to get your transcripts back.

So we\u2019ve set our sights on building the speech company. Competition from companies like Google, Amazon, and Nuance is real, but none of these approach speech recognition like we do. We've rebuilt the entire speech processing stack, replacing heuristics and stats based speech processing with fully end-to-end deep learning (we use CNNs and RNNs). Using GPUs, we train speech models to learn customer\u2019s unique vocabularies, accents, product names, and acoustic environments. This can be the difference between correctly capturing \u201cwasn\u2019t delivered\u201d and \u201cwas in the liver.\u201d We\u2019ve focused on speed since we think that\u2019s very important for exploration and scale. Our API returns hour-long transcripts interactively in seconds. It\u2019s a tool many businesses wish they had.

So far we\u2019ve released tools that:

  - transcribe speech with timestamps\n  - support real-time streaming\n  - have multi-channel support\n  - understand multiple languages (in beta now)\n  - allow you to deeply search for keywords and phrases\n  - transcribe to phonemes\n  - get more accurate with use\n
\nSome of those are better mousetraps of things you\u2019re familiar with and some are completely new levers to pull in your audio data. We\u2019ve built the core on English but now we\u2019re releasing the tools for all of the Americas. (aside: You can transfer learn speech and it works well!)

Accuracy will continue to improve for transcription, but I think we can do more. It's such a large problem, and we really want to make a dent in \u201csolving speech\u201d. That means asking, truly: \u201cWhat can a human do?\u201c

People can, with little context, jump into a conversation and determine:

  - What are the words? When are they said? Who said what?\n  - Is this person young/old? Male/Female? Exhausted/energetic?\n  - Where is there confusion?\n  - What language are they speaking? What\u2019s the speaker\u2019s accent?\n  - What\u2019s the topic of the conversation? Small talk or real? Is it going well?\n
\nSome of those things are being worked on now: additional language support, language and accent detection, sentiment analysis, auto-summarization, topic modeling, and more.

We\u2019d love to hear your feedback and ideas.","title":"Launch HN: Deepgram (YC W16) \u2013 Scalable Speech API for Businesses","updated_at":"2024-09-20T03:26:29Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"ascertain"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["deepgram"],"value":"DeepGram (YC W16) Is Building a Google for Audio"},"url":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["deepgram"],"value":"http://blog.ycombinator.com/deepgram-yc-w16-is-building-a-google-for-audio"}},"_tags":["story","author_ascertain","story_11294078"],"author":"ascertain","children":[11294238,11294485,11294779,11295632,11296218,11307894],"created_at":"2016-03-16T00:20:33Z","created_at_i":1458087633,"num_comments":28,"objectID":"11294078","points":41,"story_id":11294078,"title":"DeepGram (YC W16) Is Building a Google for Audio","updated_at":"2024-09-19T23:01:32Z","url":"http://blog.ycombinator.com/deepgram-yc-w16-is-building-a-google-for-audio"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"ablwr"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["deepgram"],"value":"Add Live Captions (Deepgram, YC W16) to Your Daily (YC W16) Prebuilt Calls"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://www.daily.co/blog/add-live-captions-to-your-daily-prebuilt-calls/"}},"_tags":["story","author_ablwr","story_31156760"],"author":"ablwr","created_at":"2022-04-25T16:02:20Z","created_at_i":1650902540,"num_comments":0,"objectID":"31156760","points":20,"story_id":31156760,"title":"Add Live Captions (Deepgram, YC W16) to Your Daily (YC W16) Prebuilt Calls","updated_at":"2024-09-20T10:59:24Z","url":"https://www.daily.co/blog/add-live-captions-to-your-daily-prebuilt-calls/"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"jasondrowley"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["deepgram"],"value":"Deepgram Completes $72M Series B Round to Build the Future of Speech AI"},"url":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["deepgram"],"value":"https://blog.deepgram.com/deepgram-72-million-series-b-defines-future-of-ai-speech-understanding/"}},"_tags":["story","author_jasondrowley","story_33786715"],"author":"jasondrowley","created_at":"2022-11-29T12:29:16Z","created_at_i":1669724956,"num_comments":0,"objectID":"33786715","points":19,"story_id":33786715,"title":"Deepgram Completes $72M Series B Round to Build the Future of Speech AI","updated_at":"2024-09-20T12:41:48Z","url":"https://blog.deepgram.com/deepgram-72-million-series-b-defines-future-of-ai-speech-understanding/"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"dsr12"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["deepgram"],"value":"Launching a \u201cGoogle for sound\u201d, Deepgram raises $1.8M"},"url":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["deepgram"],"value":"https://techcrunch.com/2016/09/27/launching-a-google-for-sound-deepgram-raises-1-8-million/"}},"_tags":["story","author_dsr12","story_12589873"],"author":"dsr12","created_at":"2016-09-27T14:37:21Z","created_at_i":1474987041,"num_comments":0,"objectID":"12589873","points":10,"story_id":12589873,"title":"Launching a \u201cGoogle for sound\u201d, Deepgram raises $1.8M","updated_at":"2024-09-19T23:47:59Z","url":"https://techcrunch.com/2016/09/27/launching-a-google-for-sound-deepgram-raises-1-8-million/"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"pstrav"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["deepgram"],"value":"Benchmarking STT providers on real calls (Deepgram 15.9% vs. OpenAI 39.8% WER)"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://twitter.com/pstrav/status/2018416957003866564"}},"_tags":["story","author_pstrav","story_46874433"],"author":"pstrav","children":[46874434,46874575,46874582],"created_at":"2026-02-03T17:53:18Z","created_at_i":1770141198,"num_comments":1,"objectID":"46874433","points":9,"story_id":46874433,"title":"Benchmarking STT providers on real calls (Deepgram 15.9% vs. OpenAI 39.8% WER)","updated_at":"2026-03-05T23:30:29Z","url":"https://twitter.com/pstrav/status/2018416957003866564"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"easwee"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["deepgram"],"value":"Show HN: Compare Speech APIs Live (OpenAI, Google, Deepgram, Soniox, etc.)"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://soniox.com/compare/"}},"_tags":["story","author_easwee","story_44569832","show_hn"],"author":"easwee","children":[44569833],"created_at":"2025-07-15T10:51:14Z","created_at_i":1752576674,"num_comments":1,"objectID":"44569832","points":7,"story_id":44569832,"title":"Show HN: Compare Speech APIs Live (OpenAI, Google, Deepgram, Soniox, etc.)","updated_at":"2025-07-19T13:43:12Z","url":"https://soniox.com/compare/"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"vhiremath4"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["deepgram"],"value":"Deepgram \u2013 Use AI to spot keywords and get insights in audio"},"url":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["deepgram"],"value":"https://www.deepgram.com/"}},"_tags":["story","author_vhiremath4","story_11307320"],"author":"vhiremath4","children":[11307341],"created_at":"2016-03-17T20:26:16Z","created_at_i":1458246376,"num_comments":1,"objectID":"11307320","points":4,"story_id":11307320,"title":"Deepgram \u2013 Use AI to spot keywords and get insights in audio","updated_at":"2024-09-19T22:55:43Z","url":"https://www.deepgram.com/"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"lukeocodes"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["deepgram"],"value":"We launched the Deepgram CLI, a command-line interface for transcription, speech synthesis, text analysis, account management, and MCP-based AI workflows.

The main idea was to make Deepgram feel native in the terminal instead of forcing people into one-off curl snippets or small throwaway apps just to test a workflow. It also features agent-aware heuristics. We looked at what fingerprint agents leave on running apps in their sandbox/environment, and decided it was better than defaulting to agent mode. If we miss any, the standard --help output has a catch-all tip for agents parsing it.

It supports local files, URLs, microphone input, stdin pipelines, structured output, named profiles, shell completion, setting up skills for your project (also found at https://github.com/deepgram/skills), and MCP server mode for AI tools.

Docs: https://developers.deepgram.com/cli/getting-started\nLaunch site: https://cli.deepgram.com\nRepo: https://github.com/deepgram/cli

Happy to answer questions about the architecture, plugin system, or MCP support."},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["deepgram"],"value":"Show HN: Deepgram releases Deepgram CLI (`dg`) an agent-aware CLI"},"url":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["deepgram"],"value":"https://cli.deepgram.com/"}},"_tags":["story","author_lukeocodes","story_47791087","show_hn"],"author":"lukeocodes","created_at":"2026-04-16T10:27:26Z","created_at_i":1776335246,"num_comments":0,"objectID":"47791087","points":4,"story_id":47791087,"story_text":"We launched the Deepgram CLI, a command-line interface for transcription, speech synthesis, text analysis, account management, and MCP-based AI workflows.

The main idea was to make Deepgram feel native in the terminal instead of forcing people into one-off curl snippets or small throwaway apps just to test a workflow. It also features agent-aware heuristics. We looked at what fingerprint agents leave on running apps in their sandbox/environment, and decided it was better than defaulting to agent mode. If we miss any, the standard --help output has a catch-all tip for agents parsing it.

It supports local files, URLs, microphone input, stdin pipelines, structured output, named profiles, shell completion, setting up skills for your project (also found at https://github.com/deepgram/skills), and MCP server mode for AI tools.

Docs: https://developers.deepgram.com/cli/getting-started\nLaunch site: https://cli.deepgram.com\nRepo: https://github.com/deepgram/cli

Happy to answer questions about the architecture, plugin system, or MCP support.","title":"Show HN: Deepgram releases Deepgram CLI (`dg`) an agent-aware CLI","updated_at":"2026-04-17T04:17:47Z","url":"https://cli.deepgram.com/"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"bcaulfield"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["deepgram"],"value":"Nvidia Invests in Speech Recognition Startup Deepgram"},"url":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["deepgram"],"value":"https://blogs.nvidia.com/blog/2018/03/26/nvidia-invests-in-speech-recognition-startup-deepgram/"}},"_tags":["story","author_bcaulfield","story_16680352"],"author":"bcaulfield","created_at":"2018-03-26T16:54:42Z","created_at_i":1522083282,"num_comments":0,"objectID":"16680352","points":4,"story_id":16680352,"title":"Nvidia Invests in Speech Recognition Startup Deepgram","updated_at":"2024-09-20T02:17:57Z","url":"https://blogs.nvidia.com/blog/2018/03/26/nvidia-invests-in-speech-recognition-startup-deepgram/"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"tester457"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["deepgram"],"value":"Deepgram Whisper Cloud: 3X Faster and 20% Cheaper Than OpenAI\u2019s"},"url":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["deepgram"],"value":"https://deepgram.com/learn/improved-whisper-api"}},"_tags":["story","author_tester457","story_36930702"],"author":"tester457","children":[36930847],"created_at":"2023-07-30T12:57:23Z","created_at_i":1690721843,"num_comments":3,"objectID":"36930702","points":3,"story_id":36930702,"title":"Deepgram Whisper Cloud: 3X Faster and 20% Cheaper Than OpenAI\u2019s","updated_at":"2024-09-20T14:45:37Z","url":"https://deepgram.com/learn/improved-whisper-api"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"mooreds"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["deepgram"],"value":"Visualizing Dungeons and Dragons with Stable Diffusion and Deepgram [video]"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://www.youtube.com/watch?v=MR6DTMGbGX0"}},"_tags":["story","author_mooreds","story_36355843"],"author":"mooreds","created_at":"2023-06-16T13:01:26Z","created_at_i":1686920486,"num_comments":0,"objectID":"36355843","points":3,"story_id":36355843,"title":"Visualizing Dungeons and Dragons with Stable Diffusion and Deepgram [video]","updated_at":"2024-09-20T14:24:56Z","url":"https://www.youtube.com/watch?v=MR6DTMGbGX0"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"innovatorved"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["deepgram"],"value":"Show HN: Self-hostable STT service with a Deepgram-compatible listen API"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://vedgupta.in/projects/whisper-api/"}},"_tags":["story","author_innovatorved","story_47732461","show_hn"],"author":"innovatorved","created_at":"2026-04-11T17:40:42Z","created_at_i":1775929242,"num_comments":0,"objectID":"47732461","points":2,"story_id":47732461,"title":"Show HN: Self-hostable STT service with a Deepgram-compatible listen API","updated_at":"2026-04-11T17:42:26Z","url":"https://vedgupta.in/projects/whisper-api/"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"domrdy"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["deepgram"],"value":"Deepgram's Voice Agent API \u2013 Deepgram"},"url":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["deepgram"],"value":"https://deepgram.com/learn/introducing-ai-voice-agent-api"}},"_tags":["story","author_domrdy","story_41594062"],"author":"domrdy","children":[41594534],"created_at":"2024-09-19T17:15:21Z","created_at_i":1726766121,"num_comments":0,"objectID":"41594062","points":2,"story_id":41594062,"title":"Deepgram's Voice Agent API \u2013 Deepgram","updated_at":"2024-09-20T17:53:31Z","url":"https://deepgram.com/learn/introducing-ai-voice-agent-api"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"deepakjois"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["deepgram"],"value":"Podscript \u2013 CLI app to transcribe podcasts using ChatGPT and Deepgram"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://github.com/deepakjois/podscript"}},"_tags":["story","author_deepakjois","story_40883319"],"author":"deepakjois","created_at":"2024-07-05T14:53:04Z","created_at_i":1720191184,"num_comments":0,"objectID":"40883319","points":2,"story_id":40883319,"title":"Podscript \u2013 CLI app to transcribe podcasts using ChatGPT and Deepgram","updated_at":"2024-09-20T17:26:44Z","url":"https://github.com/deepakjois/podscript"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"riadsila"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["deepgram"],"value":"Using LangChain, Deepgram, and Mistral 7B to Build a Video Summarization App"},"url":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["deepgram"],"value":"https://www.koyeb.com/tutorials/use-langchain-deepgram-and-mistral7b-to-build-a-youtube-video-summarization-app"}},"_tags":["story","author_riadsila","story_38301725"],"author":"riadsila","created_at":"2023-11-17T10:35:40Z","created_at_i":1700217340,"num_comments":0,"objectID":"38301725","points":2,"story_id":38301725,"title":"Using LangChain, Deepgram, and Mistral 7B to Build a Video Summarization App","updated_at":"2024-09-20T15:34:28Z","url":"https://www.koyeb.com/tutorials/use-langchain-deepgram-and-mistral7b-to-build-a-youtube-video-summarization-app"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"schappim"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["deepgram"],"value":"Show HN: CLI for text-to-speech using OpenAI/Deepgram/Elevenlabs"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://github.com/schappim/gospeak"}},"_tags":["story","author_schappim","story_46770878","show_hn"],"author":"schappim","children":[46773225],"created_at":"2026-01-26T20:14:42Z","created_at_i":1769458482,"num_comments":2,"objectID":"46770878","points":1,"story_id":46770878,"title":"Show HN: CLI for text-to-speech using OpenAI/Deepgram/Elevenlabs","updated_at":"2026-03-05T23:26:12Z","url":"https://github.com/schappim/gospeak"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"4by4by4"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["deepgram"],"value":"Deepgram raises $130M at $1.3B valuation and buys a YC AI startup"},"url":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["deepgram"],"value":"https://techcrunch.com/2026/01/13/deepgram-raises-130m-at-1-3b-valuation-and-buys-a-yc-ai-startup/"}},"_tags":["story","author_4by4by4","story_46608043"],"author":"4by4by4","created_at":"2026-01-13T21:08:21Z","created_at_i":1768338501,"num_comments":0,"objectID":"46608043","points":1,"story_id":46608043,"title":"Deepgram raises $130M at $1.3B valuation and buys a YC AI startup","updated_at":"2026-03-05T23:23:05Z","url":"https://techcrunch.com/2026/01/13/deepgram-raises-130m-at-1-3b-valuation-and-buys-a-yc-ai-startup/"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"akhilnchauhan"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["deepgram"],"value":"Hey there - I've been looking for interesting ways to stand out when applying for jobs. I'm being selective in where I work next so I'm picking a handful of companies to really go after.

Instead of just applying through the front door, I want to stand out. [1]

I used the Deepgram Voice Agent API to create a Sales Roleplay, where you have to pitch Deepgram to an AI prospect!

Fun little project. Has already drummed up interest from a couple folks.

Wish me luck!

[1] https://quarter--mile.com/Standing-Out"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["deepgram"],"value":"Show HN: I built sellmedeepgram.com to help get a job at Deepgram"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://www.sellmedeepgram.com/"}},"_tags":["story","author_akhilnchauhan","story_46319361","show_hn"],"author":"akhilnchauhan","created_at":"2025-12-18T21:55:46Z","created_at_i":1766094946,"num_comments":0,"objectID":"46319361","points":1,"story_id":46319361,"story_text":"Hey there - I've been looking for interesting ways to stand out when applying for jobs. I'm being selective in where I work next so I'm picking a handful of companies to really go after.

Instead of just applying through the front door, I want to stand out. [1]

I used the Deepgram Voice Agent API to create a Sales Roleplay, where you have to pitch Deepgram to an AI prospect!

Fun little project. Has already drummed up interest from a couple folks.

Wish me luck!

[1] https://quarter--mile.com/Standing-Out","title":"Show HN: I built sellmedeepgram.com to help get a job at Deepgram","updated_at":"2026-03-05T23:14:00Z","url":"https://www.sellmedeepgram.com/"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"koakuma-chan"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["deepgram"],"value":"Deepgram Voice Agent API Now Generally Available"},"url":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["deepgram"],"value":"https://deepgram.com/learn/voice-agent-api-generally-available"}},"_tags":["story","author_koakuma-chan","story_44291116"],"author":"koakuma-chan","created_at":"2025-06-16T16:36:06Z","created_at_i":1750091766,"num_comments":0,"objectID":"44291116","points":1,"story_id":44291116,"title":"Deepgram Voice Agent API Now Generally Available","updated_at":"2025-06-16T16:39:05Z","url":"https://deepgram.com/learn/voice-agent-api-generally-available"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"mnbbrown"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["deepgram"],"value":"Why is iOS speech to text so much better at numbers, addresses and postcodes than Deepgram, whisper, etc?

It nails them - the rest are mediocre at best. Does anybody know of a model that's accurate with this sort of data?"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["deepgram"],"value":"Ask HN: iOS STT for postcodes, addresses better than deepgram?"}},"_tags":["story","author_mnbbrown","story_42914633","ask_hn"],"author":"mnbbrown","created_at":"2025-02-03T03:32:05Z","created_at_i":1738553525,"num_comments":0,"objectID":"42914633","points":1,"story_id":42914633,"story_text":"Why is iOS speech to text so much better at numbers, addresses and postcodes than Deepgram, whisper, etc?

It nails them - the rest are mediocre at best. Does anybody know of a model that's accurate with this sort of data?","title":"Ask HN: iOS STT for postcodes, addresses better than deepgram?","updated_at":"2025-02-04T03:23:36Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"kevlened"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["deepgram"],"value":"Deepgram Aura, Conversational Text-to-Speech for Voice AI Agents"},"url":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["deepgram"],"value":"https://deepgram.com/learn/aura-text-to-speech-api-waitlist"}},"_tags":["story","author_kevlened","story_38560399"],"author":"kevlened","created_at":"2023-12-07T19:05:22Z","created_at_i":1701975922,"num_comments":0,"objectID":"38560399","points":1,"story_id":38560399,"title":"Deepgram Aura, Conversational Text-to-Speech for Voice AI Agents","updated_at":"2024-09-20T15:51:58Z","url":"https://deepgram.com/learn/aura-text-to-speech-api-waitlist"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"maxt"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["deepgram"],"value":"Deepgram \u2013 Mine your audio data"},"url":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["deepgram"],"value":"https://www.deepgram.com/"}},"_tags":["story","author_maxt","story_12847824"],"author":"maxt","created_at":"2016-11-01T17:05:00Z","created_at_i":1478019900,"num_comments":0,"objectID":"12847824","points":1,"story_id":12847824,"title":"Deepgram \u2013 Mine your audio data","updated_at":"2024-09-19T23:54:25Z","url":"https://www.deepgram.com/"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"lcolucci"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["deepgram"],"value":"Hey HN, this is Lina, Andrew, and Sidney from Lemon Slice. We\u2019ve trained a custom diffusion transformer (DiT) model that achieves video streaming at 25fps and wrapped it into a demo that allows anyone to turn a photo into a real-time, talking avatar. Here\u2019s an example conversation from co-founder Andrew: https://www.youtube.com/watch?v=CeYp5xQMFZY. Try it for yourself at: https://lemonslice.com/live.

(Btw, we used to be called Infinity AI and did a Show HN under that name last year: https://news.ycombinator.com/item?id=41467704.)

Unlike existing avatar video chat platforms like HeyGen, Tolan, or Apple Memoji filters, we do not require training custom models, rigging a character ahead of time, or having a human drive the avatar. Our tech allows users to create and immediately video-call a custom character by uploading a single image. The character image can be any style - from photorealistic to cartoons, paintings, and more.

To achieve this demo, we had to do the following (among other things! but these were the hardest):

1. Training a fast DiT model. To make our video generation fast, we had to both design a model that made the right trade-offs between speed and quality, and use standard distillation approaches. We first trained a custom video diffusion transformer (DiT) from scratch that achieves excellent lip and facial expression sync to audio. To further optimize the model for speed, we applied teacher-student distillation. The distilled model achieves 25fps video generation at 256-px resolution. Purpose-built transformer ASICs will eventually allow us to stream our video model at 4k resolution.

2. Solving the infinite video problem. Most video DiT models (Sora, Runway, Kling) generate 5-second chunks. They can iteratively extend it by another 5sec by feeding the end of the 1st chunk into the start of the 2nd in an autoregressive manner. Unfortunately the models experience quality degradation after multiple extensions due to accumulation of generation errors. We developed a temporal consistency preservation technique that maintains visual coherence across long sequences. Our technique significantly reduces artifact accumulation and allows us to generate indefinitely-long videos.

3. A complex streaming architecture with minimal latency. Enabling an end-to-end avatar zoom call requires several building blocks, including voice transcription, LLM inference, and text-to-speech generation in addition to video generation. We use Deepgram as our AI voice partner. Modal as the end-to-end compute platform. And Daily.co and Pipecat to help build a parallel processing pipeline that orchestrates everything via continuously streaming chunks. Our system achieves end-to-end latency of 3-6 seconds from user input to avatar response. Our target is <2 second latency.

More technical details here: https://lemonslice.com/live/technical-report.

Current limitations that we want to solve include: (1) enabling whole-body and background motions (we\u2019re training a next-gen model for this), (2) reducing delays and improving resolution (purpose-built ASICs will help), (3) training a model on dyadic conversations so that avatars learn to listen naturally, and (4) allowing the character to \u201csee you\u201d and respond to what they see to create a more natural and engaging conversation.

We believe that generative video will usher in a new media type centered around interactivity: TV shows, movies, ads, and online courses will stop and talk to us. Our entertainment will be a mixture of passive and active experiences depending on what we\u2019re in the mood for. Well, prediction is hard, especially about the future, but that\u2019s how we see it anyway!

We\u2019d love for you to try out the demo and let us know what you think! Post your characters and/or conversation recordings below."},"title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: Lemon Slice Live \u2013 Have a video call with a transformer model"}},"_tags":["story","author_lcolucci","story_43785044","show_hn"],"author":"lcolucci","children":[43785070,43785159,43785203,43785225,43785494,43785583,43785850,43785970,43786442,43786645,43787500,43787508,43788774,43788814,43789756,43790522,43791181,43791402,43791567,43792612,43792958,43793653,43793909,43795310,43796044,43801737,43819708],"created_at":"2025-04-24T17:10:14Z","created_at_i":1745514614,"num_comments":84,"objectID":"43785044","points":195,"story_id":43785044,"story_text":"Hey HN, this is Lina, Andrew, and Sidney from Lemon Slice. We\u2019ve trained a custom diffusion transformer (DiT) model that achieves video streaming at 25fps and wrapped it into a demo that allows anyone to turn a photo into a real-time, talking avatar. Here\u2019s an example conversation from co-founder Andrew: https://www.youtube.com/watch?v=CeYp5xQMFZY. Try it for yourself at: https://lemonslice.com/live.

(Btw, we used to be called Infinity AI and did a Show HN under that name last year: https://news.ycombinator.com/item?id=41467704.)

Unlike existing avatar video chat platforms like HeyGen, Tolan, or Apple Memoji filters, we do not require training custom models, rigging a character ahead of time, or having a human drive the avatar. Our tech allows users to create and immediately video-call a custom character by uploading a single image. The character image can be any style - from photorealistic to cartoons, paintings, and more.

To achieve this demo, we had to do the following (among other things! but these were the hardest):

1. Training a fast DiT model. To make our video generation fast, we had to both design a model that made the right trade-offs between speed and quality, and use standard distillation approaches. We first trained a custom video diffusion transformer (DiT) from scratch that achieves excellent lip and facial expression sync to audio. To further optimize the model for speed, we applied teacher-student distillation. The distilled model achieves 25fps video generation at 256-px resolution. Purpose-built transformer ASICs will eventually allow us to stream our video model at 4k resolution.

2. Solving the infinite video problem. Most video DiT models (Sora, Runway, Kling) generate 5-second chunks. They can iteratively extend it by another 5sec by feeding the end of the 1st chunk into the start of the 2nd in an autoregressive manner. Unfortunately the models experience quality degradation after multiple extensions due to accumulation of generation errors. We developed a temporal consistency preservation technique that maintains visual coherence across long sequences. Our technique significantly reduces artifact accumulation and allows us to generate indefinitely-long videos.

3. A complex streaming architecture with minimal latency. Enabling an end-to-end avatar zoom call requires several building blocks, including voice transcription, LLM inference, and text-to-speech generation in addition to video generation. We use Deepgram as our AI voice partner. Modal as the end-to-end compute platform. And Daily.co and Pipecat to help build a parallel processing pipeline that orchestrates everything via continuously streaming chunks. Our system achieves end-to-end latency of 3-6 seconds from user input to avatar response. Our target is <2 second latency.

More technical details here: https://lemonslice.com/live/technical-report.

Current limitations that we want to solve include: (1) enabling whole-body and background motions (we\u2019re training a next-gen model for this), (2) reducing delays and improving resolution (purpose-built ASICs will help), (3) training a model on dyadic conversations so that avatars learn to listen naturally, and (4) allowing the character to \u201csee you\u201d and respond to what they see to create a more natural and engaging conversation.

We believe that generative video will usher in a new media type centered around interactivity: TV shows, movies, ads, and online courses will stop and talk to us. Our entertainment will be a mixture of passive and active experiences depending on what we\u2019re in the mood for. Well, prediction is hard, especially about the future, but that\u2019s how we see it anyway!

We\u2019d love for you to try out the demo and let us know what you think! Post your characters and/or conversation recordings below.","title":"Show HN: Lemon Slice Live \u2013 Have a video call with a transformer model","updated_at":"2026-01-20T00:20:15Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"nehasuresh1904"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["deepgram"],"value":"Hi HN, we\u2019re Neha and Akash from April (https://tryapril.com). We are building an AI executive assistant to help you get through emails and manage your schedule, hands-free while you drive to work, or whenever else you prefer voice interaction.

Here's a demo: https://www.youtube.com/watch?v=ISKwEyuQQEo#t=50

...and here's a second one showing more complex use cases: https://www.youtube.com/watch?v=P8APprJ3-eY.

While driving 40 mins daily from SF to Berkeley, my inbox would flood to 30+ emails and I'd have back-to-back meetings lined up. I'd reach work already behind, then spend another hour just catching up. We figured with recent advances in voice AI, we could solve this dead time problem and start doing things on the go.

You can just speak to April and it can:

- Summarize important emails and flag what needs attention

- Send replies that I dictate (it handles the formatting and tone)

- Review my calendar and reschedule meetings on the fly

- Pull context from email threads for each meeting

- Archive/organize emails into folders

April is built using Deepgram for STT. Eleven Labs for TTS - built on top of LiveKit. We built our own MCP (Model Context Protocol) servers for Google integration, which handle auth, rate limiting, and maintain conversation context across email threads.

The most interesting part has been optimizing for lowest latency given we are a tool call heavy application. We are also trying to optimize the interruption handling and turn taking to make it feel more natural.

April is available on the AppStore (iPhone/iPad). You can try us out here https://apps.apple.com/us/app/april-ai-executive-assistant/i...

Free 3-day trial, then $14.99/mo. We'd love feedback on: Which email /calendar workflows are most painful for you? What tools beyond Gmail & Google calendar would be most valuable with a voice interface?

We\u2019ll be in the thread all day to answer any questions, share more technical details and learn what would make April most useful to you. Comment away!"},"title":{"matchLevel":"none","matchedWords":[],"value":"Launch HN: April (YC S25) \u2013 Voice AI to manage your email and calendar"}},"_tags":["story","author_nehasuresh1904","story_45015230","launch_hn"],"author":"nehasuresh1904","children":[45015615,45015682,45015850,45015858,45015926,45015982,45016120,45016308,45016539,45016727,45017022,45017151,45017240,45017572,45017597,45017647,45017807,45017823,45017890,45018103,45018417,45018766,45018915,45019462,45019636,45020783,45021013,45021081,45021308,45021608,45022155,45024070,45026273],"created_at":"2025-08-25T15:56:26Z","created_at_i":1756137386,"num_comments":95,"objectID":"45015230","points":98,"story_id":45015230,"story_text":"Hi HN, we\u2019re Neha and Akash from April (https://tryapril.com). We are building an AI executive assistant to help you get through emails and manage your schedule, hands-free while you drive to work, or whenever else you prefer voice interaction.

Here's a demo: https://www.youtube.com/watch?v=ISKwEyuQQEo#t=50

...and here's a second one showing more complex use cases: https://www.youtube.com/watch?v=P8APprJ3-eY.

While driving 40 mins daily from SF to Berkeley, my inbox would flood to 30+ emails and I'd have back-to-back meetings lined up. I'd reach work already behind, then spend another hour just catching up. We figured with recent advances in voice AI, we could solve this dead time problem and start doing things on the go.

You can just speak to April and it can:

- Summarize important emails and flag what needs attention

- Send replies that I dictate (it handles the formatting and tone)

- Review my calendar and reschedule meetings on the fly

- Pull context from email threads for each meeting

- Archive/organize emails into folders

April is built using Deepgram for STT. Eleven Labs for TTS - built on top of LiveKit. We built our own MCP (Model Context Protocol) servers for Google integration, which handle auth, rate limiting, and maintain conversation context across email threads.

The most interesting part has been optimizing for lowest latency given we are a tool call heavy application. We are also trying to optimize the interruption handling and turn taking to make it feel more natural.

April is available on the AppStore (iPhone/iPad). You can try us out here https://apps.apple.com/us/app/april-ai-executive-assistant/i...

Free 3-day trial, then $14.99/mo. We'd love feedback on: Which email /calendar workflows are most painful for you? What tools beyond Gmail & Google calendar would be most valuable with a voice interface?

We\u2019ll be in the thread all day to answer any questions, share more technical details and learn what would make April most useful to you. Comment away!","title":"Launch HN: April (YC S25) \u2013 Voice AI to manage your email and calendar","updated_at":"2026-03-05T22:30:24Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"za_mike157"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["deepgram"],"value":"TLDR: We created a personalised Andrej Karpathy tutor that can response to questions about his Youtube videos in sub 1 second responses (voice-to-voice). We do this using a voice enabled RAG agent. See later in the post for demo link, Github Repo and blog write up.

A few weeks ago we released the worlds fastest voice bot, achieving 500ms voice-to-voice response times, including a 200ms delay waiting for a user to stop speaking.

After reaching the front page of HN, we thought about how we could take this a step further based on feedback we were getting from the community. Many companies were looking for a way to implement function calling and RAG with voice interfaces while retaining a low enough latency. We couldn\u2019t find many resources about how to do this online that:

1. Allowed us to achieve sub-second voice-to-voice latency\n2. Was more flexible than existing solutions. Vapi, Retell, [Bland.ai](http://Bland.ai) are too opinionated plus since they just orchestrate API\u2019s which incur network latency at every step. See requirement above\n3. The unit economics actually work at scale.

So we decided to create a implementation of our own.

Process:

As we mentioned in our previous release, if you want to achieve response times this low you need to make everything as local as possible. So below was our setup

- Local STT: Deepgram model\n- Local Embedding model: Nomic v1.5\n- Local VectorDB: Turso\n- Local LLM: Llama 3B\n- Local TTS: Deepgram model

From our previous example, the only new components where:

- Local Embedding model: We chose Nomic Embed text v1.5 model that gave a processing time of roughly ~200ms\n- Turso offers local embedded replicas combined with edgeDB\u2019s which meant we were able to achieve 0.01 second read times. Pinecone also gave us good times of 0.043 seconds.

The above changes led us to achieve sub 1 second voice-to-voice response times

Application:

With Andrej Karpathy\u2019s announcement around [Eureka Labs](https://eurekalabs.ai/), a new AI+Education company we thought we would create our very own personalised Andrej tutor.

Listen to anyone of his Youtube lectures, as soon as your start specking, the video will pause and he will reply. Once your question has been answered you can then tell him to continue with the lecture and the video will automatically start playing.

Demo: https://educationbot.cerebrium.ai/

Blog: https://www.cerebrium.ai/blog/creating-a-realtime-rag-voice-...

Github Repo: https://github.com/CerebriumAI/examples/tree/master/19-voice...

For demo purposes:

- We used OpenAI for GPT-4-mini and embeddings (its cheaper to run on a CPU than GPU\u2019s when running demos at scale. These changes add about ~1 second to the response time\n- We used Eleven labs to clone his voice to make replies sound more realistic. This adds about 300ms to the response time.

The improvements that can be made which we would like the community to contribute to are:

- Embed the video screens as well that when you ask certain questions it can show you the relevant lecture slide for the same chuck that it got context from to answer.\n- Insert the timestamps in the vectorDB timestamps so that if a question will be answered later in the lecture he can let you know

This unlocks so many use cases in education, employee training, sales etc that it would be great to see what the community builds!"},"title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: A personalised AI tutor with < 1s voice responses"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://educationbot.cerebrium.ai/"}},"_tags":["story","author_za_mike157","story_41057030","show_hn"],"author":"za_mike157","children":[41072720,41072729,41072783,41072851,41073199,41073232,41073357,41073402,41074829,41075701],"created_at":"2024-07-24T13:41:10Z","created_at_i":1721828470,"num_comments":24,"objectID":"41057030","points":72,"story_id":41057030,"story_text":"TLDR: We created a personalised Andrej Karpathy tutor that can response to questions about his Youtube videos in sub 1 second responses (voice-to-voice). We do this using a voice enabled RAG agent. See later in the post for demo link, Github Repo and blog write up.

A few weeks ago we released the worlds fastest voice bot, achieving 500ms voice-to-voice response times, including a 200ms delay waiting for a user to stop speaking.

After reaching the front page of HN, we thought about how we could take this a step further based on feedback we were getting from the community. Many companies were looking for a way to implement function calling and RAG with voice interfaces while retaining a low enough latency. We couldn\u2019t find many resources about how to do this online that:

1. Allowed us to achieve sub-second voice-to-voice latency\n2. Was more flexible than existing solutions. Vapi, Retell, [Bland.ai](http://Bland.ai) are too opinionated plus since they just orchestrate API\u2019s which incur network latency at every step. See requirement above\n3. The unit economics actually work at scale.

So we decided to create a implementation of our own.

Process:

As we mentioned in our previous release, if you want to achieve response times this low you need to make everything as local as possible. So below was our setup

- Local STT: Deepgram model\n- Local Embedding model: Nomic v1.5\n- Local VectorDB: Turso\n- Local LLM: Llama 3B\n- Local TTS: Deepgram model

From our previous example, the only new components where:

- Local Embedding model: We chose Nomic Embed text v1.5 model that gave a processing time of roughly ~200ms\n- Turso offers local embedded replicas combined with edgeDB\u2019s which meant we were able to achieve 0.01 second read times. Pinecone also gave us good times of 0.043 seconds.

The above changes led us to achieve sub 1 second voice-to-voice response times

Application:

With Andrej Karpathy\u2019s announcement around [Eureka Labs](https://eurekalabs.ai/), a new AI+Education company we thought we would create our very own personalised Andrej tutor.

Listen to anyone of his Youtube lectures, as soon as your start specking, the video will pause and he will reply. Once your question has been answered you can then tell him to continue with the lecture and the video will automatically start playing.

Demo: https://educationbot.cerebrium.ai/

Blog: https://www.cerebrium.ai/blog/creating-a-realtime-rag-voice-...

Github Repo: https://github.com/CerebriumAI/examples/tree/master/19-voice...

For demo purposes:

- We used OpenAI for GPT-4-mini and embeddings (its cheaper to run on a CPU than GPU\u2019s when running demos at scale. These changes add about ~1 second to the response time\n- We used Eleven labs to clone his voice to make replies sound more realistic. This adds about 300ms to the response time.

The improvements that can be made which we would like the community to contribute to are:

- Embed the video screens as well that when you ask certain questions it can show you the relevant lecture slide for the same chuck that it got context from to answer.\n- Insert the timestamps in the vectorDB timestamps so that if a question will be answered later in the lecture he can let you know

This unlocks so many use cases in education, employee training, sales etc that it would be great to see what the community builds!","title":"Show HN: A personalised AI tutor with < 1s voice responses","updated_at":"2025-07-26T18:05:25Z","url":"https://educationbot.cerebrium.ai/"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"michael_y"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["deepgram"],"value":"Hi HN,\nWe\u2019re excited to share Libretto \u2013 a web-based recording and editing tool designed for creators, podcasters, and teams working remotely. If you\u2019re familiar with Descript, you\u2019ll find Libretto is similar but with unique features and a focus on simplicity.

What makes Libretto unique:

Local recording for high-quality audio/video:\nLibretto records media locally on each participant\u2019s device and uploads it in the background, ensuring pristine quality even under poor network conditions.\nThis works like double-ended recording but entirely within your browser.

Powerful yet simple editing tools:\nTranscript-based editing: Edit audio/video as if you\u2019re editing text.\nAutomatic filler word removal: Get rid of "uhs" and "ums" with one click.\nCore editing tools like splitting, trimming, and cropping are intuitive for non-professionals.

Collaborative and browser-based:\nNo downloads required \u2013 everything runs in the browser.\nTeams can work together, sharing projects and edits in real time.

Built with:

LiveKit for webRTC infrastructure, Deepgram for transcriptions, Rendley\u2019s video editing SDK\nand Liveblocks for real time shared state.

Libretto is free to try, and we\u2019d love your feedback to make it better. If you\u2019re curious, check it out.\nQuestions, thoughts, or suggestions? We\u2019re here to listen!"},"title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: Libretto \u2013 Simple recording and editing, an alternative to Descript"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://libretto.fm"}},"_tags":["story","author_michael_y","story_42149580","show_hn"],"author":"michael_y","children":[42150494,42155461,42157402,42194905],"created_at":"2024-11-15T18:34:36Z","created_at_i":1731695676,"num_comments":7,"objectID":"42149580","points":31,"story_id":42149580,"story_text":"Hi HN,\nWe\u2019re excited to share Libretto \u2013 a web-based recording and editing tool designed for creators, podcasters, and teams working remotely. If you\u2019re familiar with Descript, you\u2019ll find Libretto is similar but with unique features and a focus on simplicity.

What makes Libretto unique:

Local recording for high-quality audio/video:\nLibretto records media locally on each participant\u2019s device and uploads it in the background, ensuring pristine quality even under poor network conditions.\nThis works like double-ended recording but entirely within your browser.

Powerful yet simple editing tools:\nTranscript-based editing: Edit audio/video as if you\u2019re editing text.\nAutomatic filler word removal: Get rid of "uhs" and "ums" with one click.\nCore editing tools like splitting, trimming, and cropping are intuitive for non-professionals.

Collaborative and browser-based:\nNo downloads required \u2013 everything runs in the browser.\nTeams can work together, sharing projects and edits in real time.

Built with:

LiveKit for webRTC infrastructure, Deepgram for transcriptions, Rendley\u2019s video editing SDK\nand Liveblocks for real time shared state.

Libretto is free to try, and we\u2019d love your feedback to make it better. If you\u2019re curious, check it out.\nQuestions, thoughts, or suggestions? We\u2019re here to listen!","title":"Show HN: Libretto \u2013 Simple recording and editing, an alternative to Descript","updated_at":"2024-11-20T15:44:25Z","url":"https://libretto.fm"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"DonEsquire"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["deepgram"],"value":"Testflight Link: https://testflight.apple.com/join/8YzpXxjz

I wanted to be able to practice Spanish vocab while on my commute to work so I created YapCards. It's an iOS app focused on making flashcard studying hands-free & efficient. Like Anki, but entirely spoken.

You choose a list to practice, the app reads the prompt out loud (e.g. "What charge does a proton have?"), & you respond by voice. YapCards uses AI to evaluate your response & provide feedback.

Some key features:\n Voice-only practice (no need to look at your screen)

AI evaluation & feedback

Practice publicly available lists or upload your own excel lists. Publish your lists for others to use

Uses the SM2 algorithm to manage repetition scheduling

Tech stack includes Swift, Firebase, Deepgram, ElevenLabs, and OpenAI (trying various models)

I\u2019d love feedback from fellow learners, developers, or anyone interested in voice interfaces. Especially if you\u2019ve built language tools before \u2014 I\u2019m curious about what would make this more useful or sticky. Thoughts around voice recognition / silence detection, reducing latency, and improving AI feedback greatly appreciated!

Thanks for taking a look."},"title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: YapCards (iOS) \u2013 Voice-driven flashcards with AI feedback"}},"_tags":["story","author_DonEsquire","story_43990868","show_hn"],"author":"DonEsquire","children":[43990928,43991047,43991494,43991947,44045740],"created_at":"2025-05-15T01:14:44Z","created_at_i":1747271684,"num_comments":9,"objectID":"43990868","points":20,"story_id":43990868,"story_text":"Testflight Link: https://testflight.apple.com/join/8YzpXxjz

I wanted to be able to practice Spanish vocab while on my commute to work so I created YapCards. It's an iOS app focused on making flashcard studying hands-free & efficient. Like Anki, but entirely spoken.

You choose a list to practice, the app reads the prompt out loud (e.g. "What charge does a proton have?"), & you respond by voice. YapCards uses AI to evaluate your response & provide feedback.

Some key features:\n Voice-only practice (no need to look at your screen)

AI evaluation & feedback

Practice publicly available lists or upload your own excel lists. Publish your lists for others to use

Uses the SM2 algorithm to manage repetition scheduling

Tech stack includes Swift, Firebase, Deepgram, ElevenLabs, and OpenAI (trying various models)

I\u2019d love feedback from fellow learners, developers, or anyone interested in voice interfaces. Especially if you\u2019ve built language tools before \u2014 I\u2019m curious about what would make this more useful or sticky. Thoughts around voice recognition / silence detection, reducing latency, and improving AI feedback greatly appreciated!

Thanks for taking a look.","title":"Show HN: YapCards (iOS) \u2013 Voice-driven flashcards with AI feedback","updated_at":"2025-05-31T00:43:16Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"kodjima33"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["deepgram"],"value":"Spent 4 months and built Omi for Desktop, your life architect: It sees your screen, hears your conversations and will advise you on what to do next

Basically Cluely + Rewind + Granola + Wisprflow + ChatGPT + Claude in one app

I talk to claude/chatgpt 24/7 but I find it frustrating that i have to capture/send screenshots of my screen and that it doesn't help proactively during my work

Whenever omi sees something wrong about my workflow, it will send me a proactive notification with advice. It will also point to something I'm missing.

The hardest part was to nail proactivity - after trying 20+ similar tools I didn't find a single one with smart proactive notifications based on content on your screen. I made it look at your screen every second with 4 main prompts:

1. Is the user productive or distracted?

2. Is there anything useful to say right now?

3. is there any task to add to do later?

4. is there anything important to remember about the user?

Full stack: - Swift - Rust backend - Deepgram transcription - Claude code for messaging - GPT 5.4 summaries - Gemini for embeddings and translation

Open source, stores screenshots locally, uses Claude Code for chat. Has cloud to sync with hardware or mobile app but can be disabled in settings"},"title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: Omi \u2013 watches your screen, hears conversations, tells you what to do"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://github.com/BasedHardware/omi"}},"_tags":["story","author_kodjima33","story_47784914","show_hn"],"author":"kodjima33","children":[47784977,47785113,47785116,47785126,47785263,47788031,47788782,47788789,47788923,47789059,47789601,47790341,47790342,47791621,47798427,47801785,47802752,47808491,47815102,47817823,47822056,47826059,47832865,47842626],"created_at":"2026-04-15T20:38:21Z","created_at_i":1776285501,"num_comments":13,"objectID":"47784914","points":19,"story_id":47784914,"story_text":"Spent 4 months and built Omi for Desktop, your life architect: It sees your screen, hears your conversations and will advise you on what to do next

Basically Cluely + Rewind + Granola + Wisprflow + ChatGPT + Claude in one app

I talk to claude/chatgpt 24/7 but I find it frustrating that i have to capture/send screenshots of my screen and that it doesn't help proactively during my work

Whenever omi sees something wrong about my workflow, it will send me a proactive notification with advice. It will also point to something I'm missing.

The hardest part was to nail proactivity - after trying 20+ similar tools I didn't find a single one with smart proactive notifications based on content on your screen. I made it look at your screen every second with 4 main prompts:

1. Is the user productive or distracted?

2. Is there anything useful to say right now?

3. is there any task to add to do later?

4. is there anything important to remember about the user?

Full stack: - Swift - Rust backend - Deepgram transcription - Claude code for messaging - GPT 5.4 summaries - Gemini for embeddings and translation

Open source, stores screenshots locally, uses Claude Code for chat. Has cloud to sync with hardware or mobile app but can be disabled in settings","title":"Show HN: Omi \u2013 watches your screen, hears conversations, tells you what to do","updated_at":"2026-05-03T20:22:32Z","url":"https://github.com/BasedHardware/omi"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"jmtame"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["deepgram"],"value":"Hey HN! We\u2019re Jared, Shreyas, and Varun the creators of TechInterviewer. We\u2019re building a product for software engineers to go through an entirely simulated systems design interview. Our AI interviewer, Steve, gives you a prompt and you talk out loud and draw on a whiteboard while Steve guides you through the interview and gives real-time feedback. Check out our demo:

https://app.techinterviewer.ai

Every software engineer today has to prepare for systems design interviews and have two awful options: pay hundreds of dollars for a single session with a FAANG engineer or follow silently alongside a YouTube playlist. Because there is no instant feedback while practicing, engineers often learn about their most important knowledge gaps during the course of the interview loop.

Jared and Shreyas are both senior engineers who have spent 1000s of hours preparing for and administering systems design interviews. Shreyas was an early engineer at Deepgram and spent many years tracking developments in the TTS (text to speech) space. He realized that voice interviews had potential to change the candidate experience when he starting using chatGPT to prepare for interviewing founding engineer candidates at his startup.

We\u2019re hoping that having easy access to interview feedback will level the playing field of software engineers at different skill levels. We\u2019re really excited to share this with you all and we\u2019d love any thoughts, feedback, and comments"},"title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: We built a multimodal AI interviewer for mock system design interviews"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://app.techinterviewer.ai?source=hn"}},"_tags":["story","author_jmtame","story_39055841","show_hn"],"author":"jmtame","children":[39057122,39059507,39064759,39220603],"created_at":"2024-01-19T14:29:27Z","created_at_i":1705674567,"num_comments":7,"objectID":"39055841","points":11,"story_id":39055841,"story_text":"Hey HN! We\u2019re Jared, Shreyas, and Varun the creators of TechInterviewer. We\u2019re building a product for software engineers to go through an entirely simulated systems design interview. Our AI interviewer, Steve, gives you a prompt and you talk out loud and draw on a whiteboard while Steve guides you through the interview and gives real-time feedback. Check out our demo:

https://app.techinterviewer.ai

Every software engineer today has to prepare for systems design interviews and have two awful options: pay hundreds of dollars for a single session with a FAANG engineer or follow silently alongside a YouTube playlist. Because there is no instant feedback while practicing, engineers often learn about their most important knowledge gaps during the course of the interview loop.

Jared and Shreyas are both senior engineers who have spent 1000s of hours preparing for and administering systems design interviews. Shreyas was an early engineer at Deepgram and spent many years tracking developments in the TTS (text to speech) space. He realized that voice interviews had potential to change the candidate experience when he starting using chatGPT to prepare for interviewing founding engineer candidates at his startup.

We\u2019re hoping that having easy access to interview feedback will level the playing field of software engineers at different skill levels. We\u2019re really excited to share this with you all and we\u2019d love any thoughts, feedback, and comments","title":"Show HN: We built a multimodal AI interviewer for mock system design interviews","updated_at":"2024-11-11T16:11:21Z","url":"https://app.techinterviewer.ai?source=hn"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"pritesh1908"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["deepgram"],"value":"We kept hitting the same wall building voice AI systems. Pipecat and LiveKit are great projects, genuinely. But getting it to production took us weeks of plumbing - wiring things together, handling barge-ins, setting up telephony, Knowledge base, tool calls, handling barge in etc. And every time we needed to tweak agent behavior, you were back in the code and redeploying. We just wanted to change a prompt and test it in 30 seconds. Thats why Vapi retell etc exist.

So we wrote the entire code and open sourced it as a Visual drag-and-drop for voice agents ( same as vapi or n8n for voice). Built on a Pipecat fork and BSD-2, no strings attached.\nTool calls, knowledge base, variable extraction, voicemail detection, call transfer to humans, multilingual support, post-call QA, background noise suppression, and a website widget are all included. You're not paying per-minute fees to a middleman wrapping the same APIs you'd call directly.

You can set it up with a simple docker command. It comes pre-wired with Deepgram, Cartesia, OpenAI , Speechmatics Sarvam for STT, same for TTS, and OpenAI, Gemini, groq, Openrouter, Azure on the LLM side. \nTelephony works out of the box with Twilio, Vonage , CLoudonix and Asterisk for both inbound and outbound.

There's a hosted version at app.dograh.com if self-hosting isn't your thing.

Repo: github.com/dograh-hq/dograh\nVideo walkthrough: https://youtu.be/sxiSp4JXqws

We built this out of frustration, not a thesis.\nThe tool is free to use and fully open source (and will always remain so), happy to answer questions about the data or how we built it."},"title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: We open sourced Vapi \u2013 UI included"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://github.com/dograh-hq/dograh"}},"_tags":["story","author_pritesh1908","story_47351652","show_hn"],"author":"pritesh1908","children":[47351697,47353671,47354022,47358931,47361253,47408779],"created_at":"2026-03-12T15:03:19Z","created_at_i":1773327799,"num_comments":7,"objectID":"47351652","points":8,"story_id":47351652,"story_text":"We kept hitting the same wall building voice AI systems. Pipecat and LiveKit are great projects, genuinely. But getting it to production took us weeks of plumbing - wiring things together, handling barge-ins, setting up telephony, Knowledge base, tool calls, handling barge in etc. And every time we needed to tweak agent behavior, you were back in the code and redeploying. We just wanted to change a prompt and test it in 30 seconds. Thats why Vapi retell etc exist.

So we wrote the entire code and open sourced it as a Visual drag-and-drop for voice agents ( same as vapi or n8n for voice). Built on a Pipecat fork and BSD-2, no strings attached.\nTool calls, knowledge base, variable extraction, voicemail detection, call transfer to humans, multilingual support, post-call QA, background noise suppression, and a website widget are all included. You're not paying per-minute fees to a middleman wrapping the same APIs you'd call directly.

You can set it up with a simple docker command. It comes pre-wired with Deepgram, Cartesia, OpenAI , Speechmatics Sarvam for STT, same for TTS, and OpenAI, Gemini, groq, Openrouter, Azure on the LLM side. \nTelephony works out of the box with Twilio, Vonage , CLoudonix and Asterisk for both inbound and outbound.

There's a hosted version at app.dograh.com if self-hosting isn't your thing.

Repo: github.com/dograh-hq/dograh\nVideo walkthrough: https://youtu.be/sxiSp4JXqws

We built this out of frustration, not a thesis.\nThe tool is free to use and fully open source (and will always remain so), happy to answer questions about the data or how we built it.","title":"Show HN: We open sourced Vapi \u2013 UI included","updated_at":"2026-04-03T12:06:51Z","url":"https://github.com/dograh-hq/dograh"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"jmtame"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["deepgram"],"value":"Hey HN! We're Jared, Shreyas, and Varun the creators of TechInterviewer. We're building a product for software engineers to go through an entirely simulated systems design interview. Our AI interviewer, Steve, gives you a prompt and you talk out loud and draw on a whiteboard while Steve guides you through the interview and gives real-time feedback. Check out our demo:

https://app.techinterviewer.ai

Every software engineer today has to prepare for systems design interviews and have two awful options: pay hundreds of dollars for a single session with a FAANG engineer or follow silently alongside a YouTube playlist. Because there is no instant feedback while practicing, engineers often learn about their most important knowledge gaps during the course of the interview loop.

Jared and Shreyas are both senior engineers who have spent 100s of hours preparing for and administering systems design interviews. Shreyas was an early engineer at Deepgram and spent many years tracking developments in the TTS (text to speech) space. He realized that voice interviews had potential to change the candidate experience when he starting using chatGPT to prepare for interviewing founding engineer candidates at his startup.

We're hoping that having easy access to interview feedback will level the playing field of software engineers at different skill levels. We're really excited to share this with you all and we'd love any thoughts, feedback, and comments"},"title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: We built an AI multimodal interviewer for mock system design interviews"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://app.techinterviewer.ai/"}},"_tags":["story","author_jmtame","story_39055609","show_hn"],"author":"jmtame","children":[39057383],"created_at":"2024-01-19T14:10:12Z","created_at_i":1705673412,"num_comments":1,"objectID":"39055609","points":8,"story_id":39055609,"story_text":"Hey HN! We're Jared, Shreyas, and Varun the creators of TechInterviewer. We're building a product for software engineers to go through an entirely simulated systems design interview. Our AI interviewer, Steve, gives you a prompt and you talk out loud and draw on a whiteboard while Steve guides you through the interview and gives real-time feedback. Check out our demo:

https://app.techinterviewer.ai

Every software engineer today has to prepare for systems design interviews and have two awful options: pay hundreds of dollars for a single session with a FAANG engineer or follow silently alongside a YouTube playlist. Because there is no instant feedback while practicing, engineers often learn about their most important knowledge gaps during the course of the interview loop.

Jared and Shreyas are both senior engineers who have spent 100s of hours preparing for and administering systems design interviews. Shreyas was an early engineer at Deepgram and spent many years tracking developments in the TTS (text to speech) space. He realized that voice interviews had potential to change the candidate experience when he starting using chatGPT to prepare for interviewing founding engineer candidates at his startup.

We're hoping that having easy access to interview feedback will level the playing field of software engineers at different skill levels. We're really excited to share this with you all and we'd love any thoughts, feedback, and comments","title":"Show HN: We built an AI multimodal interviewer for mock system design interviews","updated_at":"2024-09-20T16:14:02Z","url":"https://app.techinterviewer.ai/"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"hkjarral"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["deepgram"],"value":"Hi HN, I'm the creator of AVA - AI Voice Agent for Asterisk

My repo was shared here once before by someone else so I wanted to follow up with the progress since then.

https://news.ycombinator.com/item?id=46380399

I've been working with Asterisk/FreePBX systems for years. I wanted to add AI voice capabilities to legacy phone systems without paying per-minute SaaS fees or ripping out the entire telephony stack.

So I built AVA, a self-hosted AI voice agent that can integrate into any traditional phone system. While most solutions demand expensive migrations to cloud-only providers, AVA provides a self-hosted path to connect AI agents to existing phone systems while ensuring data privacy and lowering operational costs

AVA is a Dockerized Python app that sits alongside your Asterisk server. It connects via ARI (Asterisk REST Interface) and routes call audio to AI providers \u2014 OpenAI Realtime, Deepgram, Google Live API, ElevenLabs, Telnyx, or fully local models (Vosk + llama.cpp + Piper). You can mix and match STT/LLM/TTS in a modular pipeline, or use a single provider end-to-end.

Two audio transport paths: We support both AudioSocket (low-latency TCP with TLV framing) and ExternalMedia RTP (UDP, better for NAT). A transport orchestrator auto-negotiates sample rates and codecs between what Asterisk sends on the wire and what each AI provider expects \u2014 so you can run 8kHz ulaw from Asterisk into a provider that wants 24kHz linear16 without manual config.

Session lifecycle: A typed session store tracks every call from StasisStart through hangup \u2014 audio diagnostics, barge-in counts, provider state, conversation turns. Every call is fully observable and debuggable after the fact.

Barge-in and VAD were the hardest problems. We use a dual-mode VAD \u2014 WebRTC VAD combined with energy-based RMS detection, scored into a single confidence value (40% WebRTC weight, 40% energy ratio, 20% agreement bonus). Frame smoothing prevents single-frame glitches from triggering false interrupts. When barge-in fires, we kill active playback (both streaming and file-based) via ARI, flush provider audio buffers, release conversation gating tokens, and optionally suppress provider output for a configurable window to prevent pre-barge audio from re-queuing. The system supports three interrupt sources: local VAD, Asterisk's native talk detection events, and provider-side interruption signals.

The hardest latency challenge was bridging legacy SIP/RTP with modern WebSocket streams. We use a two-container architecture: a lightweight orchestrator for ARI state management and an optional heavier container for local model inference. There are 6 pre-validated golden baseline configs if you just want something working out of the box, plus an Admin UI for visual setup.

Try the live demo: (925)-736-6718 Option 5 for Google, 6 for Deepgram, 7 for Openai realtime, 8 for Local hybrid and 9 for Elevenlabs

Code is MIT. I'd love feedback on the transport layer (src/core/transport_orchestrator.py) and the VAD tuning (src/core/vad_manager.py)."},"title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: Ava \u2013 AI Voice Agent for Traditional Phone Systems(Python+Asterisk/ARI)"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://github.com/hkjarral/AVA-AI-Voice-Agent-for-Asterisk"}},"_tags":["story","author_hkjarral","story_47354996","show_hn"],"author":"hkjarral","children":[47360696,47472391],"created_at":"2026-03-12T18:16:24Z","created_at_i":1773339384,"num_comments":1,"objectID":"47354996","points":7,"story_id":47354996,"story_text":"Hi HN, I'm the creator of AVA - AI Voice Agent for Asterisk

My repo was shared here once before by someone else so I wanted to follow up with the progress since then.

https://news.ycombinator.com/item?id=46380399

I've been working with Asterisk/FreePBX systems for years. I wanted to add AI voice capabilities to legacy phone systems without paying per-minute SaaS fees or ripping out the entire telephony stack.

So I built AVA, a self-hosted AI voice agent that can integrate into any traditional phone system. While most solutions demand expensive migrations to cloud-only providers, AVA provides a self-hosted path to connect AI agents to existing phone systems while ensuring data privacy and lowering operational costs

AVA is a Dockerized Python app that sits alongside your Asterisk server. It connects via ARI (Asterisk REST Interface) and routes call audio to AI providers \u2014 OpenAI Realtime, Deepgram, Google Live API, ElevenLabs, Telnyx, or fully local models (Vosk + llama.cpp + Piper). You can mix and match STT/LLM/TTS in a modular pipeline, or use a single provider end-to-end.

Two audio transport paths: We support both AudioSocket (low-latency TCP with TLV framing) and ExternalMedia RTP (UDP, better for NAT). A transport orchestrator auto-negotiates sample rates and codecs between what Asterisk sends on the wire and what each AI provider expects \u2014 so you can run 8kHz ulaw from Asterisk into a provider that wants 24kHz linear16 without manual config.

Session lifecycle: A typed session store tracks every call from StasisStart through hangup \u2014 audio diagnostics, barge-in counts, provider state, conversation turns. Every call is fully observable and debuggable after the fact.

Barge-in and VAD were the hardest problems. We use a dual-mode VAD \u2014 WebRTC VAD combined with energy-based RMS detection, scored into a single confidence value (40% WebRTC weight, 40% energy ratio, 20% agreement bonus). Frame smoothing prevents single-frame glitches from triggering false interrupts. When barge-in fires, we kill active playback (both streaming and file-based) via ARI, flush provider audio buffers, release conversation gating tokens, and optionally suppress provider output for a configurable window to prevent pre-barge audio from re-queuing. The system supports three interrupt sources: local VAD, Asterisk's native talk detection events, and provider-side interruption signals.

The hardest latency challenge was bridging legacy SIP/RTP with modern WebSocket streams. We use a two-container architecture: a lightweight orchestrator for ARI state management and an optional heavier container for local model inference. There are 6 pre-validated golden baseline configs if you just want something working out of the box, plus an Admin UI for visual setup.

Try the live demo: (925)-736-6718 Option 5 for Google, 6 for Deepgram, 7 for Openai realtime, 8 for Local hybrid and 9 for Elevenlabs

Code is MIT. I'd love feedback on the transport layer (src/core/transport_orchestrator.py) and the VAD tuning (src/core/vad_manager.py).","title":"Show HN: Ava \u2013 AI Voice Agent for Traditional Phone Systems(Python+Asterisk/ARI)","updated_at":"2026-03-23T04:30:19Z","url":"https://github.com/hkjarral/AVA-AI-Voice-Agent-for-Asterisk"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"sgoudelis"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["deepgram"],"value":"Hello Hacker News community,

I'm excited to share my open-source project, Ground Station \u2013 a full-featured software solution for satellite tracking and radio communication. Designed for amateur radio operators, satellite enthusiasts, and researchers, it provides a comprehensive and easy-to-use platform.

Key Features include:

- Real-time Satellite Tracking: Precise orbital models and automatic TLE updates.

- SDR Integration: Stream and record live radio signals from a wide range of SDR devices (RTL-SDR, SoapySDR, UHD/USRP).

- AI-Powered Transcription: Real-time speech-to-text for demodulated audio via Gemini Live or Deepgram.

- Automated Observations: Schedule and execute satellite passes without manual intervention, including IQ/audio recording and protocol decoding (SSTV, FSK, GFSK, GMSK, BPSK).

- SatDump Integration: Decode weather satellite images (METEOR-M2 LRPT/HRPT).

- Responsive Web Interface: Control everything from desktop, tablet, or mobile.

- Automated Antenna Rotator & Rig Control: Interface with popular hardware with Doppler correction.

This project was built with the help of Claude Code and Codex.

You can find the project here: https://github.com/sgoudelis/ground-station

I'd appreciate any feedback, suggestions, or contributions from the community!"},"title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: Ground Station \u2013 All-in-one satellite monitoring suite (Python, SDR)"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://github.com/sgoudelis/ground-station"}},"_tags":["story","author_sgoudelis","story_47006167","show_hn"],"author":"sgoudelis","children":[47006385],"created_at":"2026-02-13T18:45:28Z","created_at_i":1771008328,"num_comments":1,"objectID":"47006167","points":7,"story_id":47006167,"story_text":"Hello Hacker News community,

I'm excited to share my open-source project, Ground Station \u2013 a full-featured software solution for satellite tracking and radio communication. Designed for amateur radio operators, satellite enthusiasts, and researchers, it provides a comprehensive and easy-to-use platform.

Key Features include:

- Real-time Satellite Tracking: Precise orbital models and automatic TLE updates.

- SDR Integration: Stream and record live radio signals from a wide range of SDR devices (RTL-SDR, SoapySDR, UHD/USRP).

- AI-Powered Transcription: Real-time speech-to-text for demodulated audio via Gemini Live or Deepgram.

- Automated Observations: Schedule and execute satellite passes without manual intervention, including IQ/audio recording and protocol decoding (SSTV, FSK, GFSK, GMSK, BPSK).

- SatDump Integration: Decode weather satellite images (METEOR-M2 LRPT/HRPT).

- Responsive Web Interface: Control everything from desktop, tablet, or mobile.

- Automated Antenna Rotator & Rig Control: Interface with popular hardware with Doppler correction.

This project was built with the help of Claude Code and Codex.

You can find the project here: https://github.com/sgoudelis/ground-station

I'd appreciate any feedback, suggestions, or contributions from the community!","title":"Show HN: Ground Station \u2013 All-in-one satellite monitoring suite (Python, SDR)","updated_at":"2026-03-05T23:32:05Z","url":"https://github.com/sgoudelis/ground-station"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"angelos_7"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["deepgram"],"value":"I've built an interview assistant that sits invisibly on your screen during technical interviews. \nIt can:

- Take screenshots of coding problems with Cmd+H\n- Listen to questions through voice transcription (using Deepgram) with Cmd+E\n- Generate complete solutions using GPT-4 Vision and GPT-4.1\n- Stay nearly invisible with transparency and hotkeys \n- Works on all meeting and screen-recording software.\n- Buy once, no subscription, yours forever. All updates are included. Email Support

It's built with Electron, React, and OpenAI's API to fetch answers and Deepgram s.

The app is designed for remote technical interviews where you need help but can't have obvious assistance visible on your screen.

Looking for feedback on the concept and implementation. Would love to hear your thoughts!

Try for Free. \nBuy it if you like with coupon code SHADOWLAUNCH50 for 50% off."},"title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: Shadow Interview Assistant \u2013 AI that sees and listens your interviews"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://interview-assistant-ai.com/shadow-interview-assistant.html"}},"_tags":["story","author_angelos_7","story_43690842","show_hn"],"author":"angelos_7","created_at":"2025-04-15T09:54:50Z","created_at_i":1744710890,"num_comments":0,"objectID":"43690842","points":6,"story_id":43690842,"story_text":"I've built an interview assistant that sits invisibly on your screen during technical interviews. \nIt can:

- Take screenshots of coding problems with Cmd+H\n- Listen to questions through voice transcription (using Deepgram) with Cmd+E\n- Generate complete solutions using GPT-4 Vision and GPT-4.1\n- Stay nearly invisible with transparency and hotkeys \n- Works on all meeting and screen-recording software.\n- Buy once, no subscription, yours forever. All updates are included. Email Support

It's built with Electron, React, and OpenAI's API to fetch answers and Deepgram s.

The app is designed for remote technical interviews where you need help but can't have obvious assistance visible on your screen.

Looking for feedback on the concept and implementation. Would love to hear your thoughts!

Try for Free. \nBuy it if you like with coupon code SHADOWLAUNCH50 for 50% off.","title":"Show HN: Shadow Interview Assistant \u2013 AI that sees and listens your interviews","updated_at":"2025-04-15T10:23:56Z","url":"https://interview-assistant-ai.com/shadow-interview-assistant.html"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"arbayi"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["deepgram"],"value":"For a fun side weekend project we gave our own OpenClaw agent called "Mahmut" a face and had a live interview with it. It went surprisingly well.

Here's a sneak peek from the interview: https://x.com/ptservlor/status/2024597444890128767

User speaks, Deepgram transcribes it, OpenClaw Gateway routes it to your agent, ElevenLabs turns the response into speech, and LemonSlice generates a lip synced avatar from the audio. Everything streams over LiveKit in real time.

Latency is about 1 to 2 seconds end to end depending on the LLM. The lip sync from LemonSlice honestly surprised us, it works way better than we expected.

The skill repo has a complete Python example, env setup, troubleshooting guide, and a Next.js frontend guide if you want to build your own web UI for it."},"title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: Give your OpenClaw agent a face and voice with LiveKit and LemonSlice"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://github.com/openserv-labs/openclaw-voice-avatar"}},"_tags":["story","author_arbayi","story_47096208","show_hn"],"author":"arbayi","created_at":"2026-02-21T00:50:20Z","created_at_i":1771635020,"num_comments":0,"objectID":"47096208","points":5,"story_id":47096208,"story_text":"For a fun side weekend project we gave our own OpenClaw agent called "Mahmut" a face and had a live interview with it. It went surprisingly well.

Here's a sneak peek from the interview: https://x.com/ptservlor/status/2024597444890128767

User speaks, Deepgram transcribes it, OpenClaw Gateway routes it to your agent, ElevenLabs turns the response into speech, and LemonSlice generates a lip synced avatar from the audio. Everything streams over LiveKit in real time.

Latency is about 1 to 2 seconds end to end depending on the LLM. The lip sync from LemonSlice honestly surprised us, it works way better than we expected.

The skill repo has a complete Python example, env setup, troubleshooting guide, and a Next.js frontend guide if you want to build your own web UI for it.","title":"Show HN: Give your OpenClaw agent a face and voice with LiveKit and LemonSlice","updated_at":"2026-03-05T23:36:51Z","url":"https://github.com/openserv-labs/openclaw-voice-avatar"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"kamiljk"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["deepgram"],"value":"Hey HN, I\u2019m one of two people working on Hyperscreen AI, an AI that does mock technical interviews with you so you can pass more tech screens.

I shut my YC startup down a year ago and had to look for a job afterward. Even though I had interned at and landed multiple FAANG offers out of college, I spent almost 6 months looking for a job after my company failed. I was fortunate enough to get technical screens due to my network, but I failed many of them because I didn\u2019t have anybody to practice with. All of my friends either already had jobs and weren\u2019t looking, or were still working on their startups. It was incredibly lonely to grind leetcode in my parents\u2019 house for 8+ hours a day, 6 days a week, for months.

I eventually got fed up with not having anyone to practice interviewing with, so I built a hacky voice bot on top of the latest LLMs to practice with. I was finally able to land a SWE job at a great late-stage startup.

If someone with my background struggled this much to land a job, I can\u2019t imagine how difficult it must be for many of the people who aren\u2019t as privileged and have fewer interview opportunities to begin with. I went to a Tier 3 college when the market was still good, and, in my best year, I was only able to get phone screens with 4 companies. With so few chances, you need to be perfect and every edge you can get while prepping helps.

My friend and I went to a hackathon a few months ago to build a more fleshed-out version of the tool I built for myself. After a few rounds of iteration with initial customers, it\u2019s finally ready.

Now for what you\u2019ve all been waiting for, the juicy technical details. We\u2019re using a simple React front end and Flask backend, with a series of model vendors provided to us through Vapi to handle the model audio interaction (OpenAI TTS, Deepgram STT). We\u2019re using OpenAI for the feedback. The model gets access to your code via us sending it any time you speak.

Others are building similar products, but none of them have gained substantial brand awareness, and none were (in my opinion) as meticulously designed or thought through as ours is.

We\u2019ve gotten feedback that this is just another AI wrapper, but we don\u2019t think that\u2019s true. You can\u2019t effectively run a technical interview for free using just ChatGPT \u2014 we\u2019ve tried it. It gives away answers, overexplains, and struggles to guide you. Not to mention the poorer UX, since you can\u2019t easily sync your code while speaking to it. The way to get around this would be building a similar product to us, but then you\u2019d have to pay for the audio streaming service yourself (either OpenAI real time, which is still in beta and very expensive, or something like Vapi which you\u2019d still have to pay for). In other words, you need to pay if you want to run a mock technical interview.

We\u2019re still experimenting with our business model (as most new businesses tend to), but for now, we make money by charging a subscription for a fixed amount of time that our customers get each month. Currently, that\u2019s $9.75 for an hour of interview time per month. My co-founder\u2019s mom has paid upwards of $150 for 45-minute interviews of varying quality. Meanwhile, we currently charge more than 10x less for the same amount of time with a more consistent experience and detailed feedback. You can find them cheaper than this, but we have yet to see a service where you can interview with a human for cheaper than $9.75/hour. If you have friends who can interview you for free, that\u2019s great!

To be honest, writing this post is pretty surreal. I never did a Show HN for my last startup and held myself back from even working on this project or with a cofounder for a long time because I was so traumatized by my last experience. In retrospect, I\u2019m glad I did it, and I\u2019m thrilled to share what we\u2019ve built so far.

We\u2019re really excited to hear all of your feedback, whether or not you decide to try it!"},"title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: I built an AI to mock technical interview me after I shut down my YC co"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://www.hyperscreen.ai/"}},"_tags":["story","author_kamiljk","story_43081613","show_hn"],"author":"kamiljk","created_at":"2025-02-17T17:59:27Z","created_at_i":1739815167,"num_comments":0,"objectID":"43081613","points":5,"story_id":43081613,"story_text":"Hey HN, I\u2019m one of two people working on Hyperscreen AI, an AI that does mock technical interviews with you so you can pass more tech screens.

I shut my YC startup down a year ago and had to look for a job afterward. Even though I had interned at and landed multiple FAANG offers out of college, I spent almost 6 months looking for a job after my company failed. I was fortunate enough to get technical screens due to my network, but I failed many of them because I didn\u2019t have anybody to practice with. All of my friends either already had jobs and weren\u2019t looking, or were still working on their startups. It was incredibly lonely to grind leetcode in my parents\u2019 house for 8+ hours a day, 6 days a week, for months.

I eventually got fed up with not having anyone to practice interviewing with, so I built a hacky voice bot on top of the latest LLMs to practice with. I was finally able to land a SWE job at a great late-stage startup.

If someone with my background struggled this much to land a job, I can\u2019t imagine how difficult it must be for many of the people who aren\u2019t as privileged and have fewer interview opportunities to begin with. I went to a Tier 3 college when the market was still good, and, in my best year, I was only able to get phone screens with 4 companies. With so few chances, you need to be perfect and every edge you can get while prepping helps.

My friend and I went to a hackathon a few months ago to build a more fleshed-out version of the tool I built for myself. After a few rounds of iteration with initial customers, it\u2019s finally ready.

Now for what you\u2019ve all been waiting for, the juicy technical details. We\u2019re using a simple React front end and Flask backend, with a series of model vendors provided to us through Vapi to handle the model audio interaction (OpenAI TTS, Deepgram STT). We\u2019re using OpenAI for the feedback. The model gets access to your code via us sending it any time you speak.

Others are building similar products, but none of them have gained substantial brand awareness, and none were (in my opinion) as meticulously designed or thought through as ours is.

We\u2019ve gotten feedback that this is just another AI wrapper, but we don\u2019t think that\u2019s true. You can\u2019t effectively run a technical interview for free using just ChatGPT \u2014 we\u2019ve tried it. It gives away answers, overexplains, and struggles to guide you. Not to mention the poorer UX, since you can\u2019t easily sync your code while speaking to it. The way to get around this would be building a similar product to us, but then you\u2019d have to pay for the audio streaming service yourself (either OpenAI real time, which is still in beta and very expensive, or something like Vapi which you\u2019d still have to pay for). In other words, you need to pay if you want to run a mock technical interview.

We\u2019re still experimenting with our business model (as most new businesses tend to), but for now, we make money by charging a subscription for a fixed amount of time that our customers get each month. Currently, that\u2019s $9.75 for an hour of interview time per month. My co-founder\u2019s mom has paid upwards of $150 for 45-minute interviews of varying quality. Meanwhile, we currently charge more than 10x less for the same amount of time with a more consistent experience and detailed feedback. You can find them cheaper than this, but we have yet to see a service where you can interview with a human for cheaper than $9.75/hour. If you have friends who can interview you for free, that\u2019s great!

To be honest, writing this post is pretty surreal. I never did a Show HN for my last startup and held myself back from even working on this project or with a cofounder for a long time because I was so traumatized by my last experience. In retrospect, I\u2019m glad I did it, and I\u2019m thrilled to share what we\u2019ve built so far.

We\u2019re really excited to hear all of your feedback, whether or not you decide to try it!","title":"Show HN: I built an AI to mock technical interview me after I shut down my YC co","updated_at":"2025-02-17T19:35:50Z","url":"https://www.hyperscreen.ai/"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"mdwelsh"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["deepgram"],"value":"Hey HN, I\u2019m excited to show off a project I have been building the last few months called Podverse \u2014 it automatically generates transcripts, summaries, and a chatbot from a podcast\u2019s RSS feed, and makes it easy to embed AI features onto any website.

You can check it out at https://podverse.ai.

See it in action on an example podcast here: https://www.podverse.ai/podcast/omnibus

I\u2019m a former engineer at Google and Apple, former CS prof at Harvard, and was a co-founder at Fixie.ai. I truly feel that LLMs are a revolution in our field, and I wanted to bring something to life that bridged between two deep interests of mine \u2014 AI and podcasts.

I\u2019ve been listening a ton of podcasts lately, but it's frustrating to find relevant episodes, or even content that I had listened to previously. I also know that podcasters want to take advantage of the latest AI tech to connect with and expand their audience, but they're not usually equipped to bring all the pieces together. I wanted to build a complete, end-to-end solution for podcasters that is easy to use.

All you have to do is provide your RSS feed URL, and Podverse does the rest \u2014 automatically ingesting new episodes as you release them, generating transcripts and summaries, identifying speakers, and populating a custom chatbot that knows about all of the content of your podcast.

I built the site from scratch over the last few months (in my copious spare time!). It\u2019s built entirely using Next.js deployed on Vercel \u2014 no backend servers beyond Vercel\u2019s server components and API handlers. For running background tasks, like podcast ingestion and transcription, I used Inngest, which fills a big gap in the Next.js stack. Deepgram is used for ASR, which is fast, affordable, and very high quality. Supabase handles the storage, including the vector database (using pgvector). Auth is done with Clerk, which was super easy to integrate. Happy to answer any questions about the stack.

I\u2019d love to get your feedback, and especially hear from podcasters to learn more about what features you\u2019d like to see. Thanks!"},"title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: Podverse \u2013 AI Superpowers for Podcasters"}},"_tags":["story","author_mdwelsh","story_41697569","show_hn"],"author":"mdwelsh","children":[41703929],"created_at":"2024-09-30T14:33:10Z","created_at_i":1727706790,"num_comments":1,"objectID":"41697569","points":4,"story_id":41697569,"story_text":"Hey HN, I\u2019m excited to show off a project I have been building the last few months called Podverse \u2014 it automatically generates transcripts, summaries, and a chatbot from a podcast\u2019s RSS feed, and makes it easy to embed AI features onto any website.

You can check it out at https://podverse.ai.

See it in action on an example podcast here: https://www.podverse.ai/podcast/omnibus

I\u2019m a former engineer at Google and Apple, former CS prof at Harvard, and was a co-founder at Fixie.ai. I truly feel that LLMs are a revolution in our field, and I wanted to bring something to life that bridged between two deep interests of mine \u2014 AI and podcasts.

I\u2019ve been listening a ton of podcasts lately, but it's frustrating to find relevant episodes, or even content that I had listened to previously. I also know that podcasters want to take advantage of the latest AI tech to connect with and expand their audience, but they're not usually equipped to bring all the pieces together. I wanted to build a complete, end-to-end solution for podcasters that is easy to use.

All you have to do is provide your RSS feed URL, and Podverse does the rest \u2014 automatically ingesting new episodes as you release them, generating transcripts and summaries, identifying speakers, and populating a custom chatbot that knows about all of the content of your podcast.

I built the site from scratch over the last few months (in my copious spare time!). It\u2019s built entirely using Next.js deployed on Vercel \u2014 no backend servers beyond Vercel\u2019s server components and API handlers. For running background tasks, like podcast ingestion and transcription, I used Inngest, which fills a big gap in the Next.js stack. Deepgram is used for ASR, which is fast, affordable, and very high quality. Supabase handles the storage, including the vector database (using pgvector). Auth is done with Clerk, which was super easy to integrate. Happy to answer any questions about the stack.

I\u2019d love to get your feedback, and especially hear from podcasters to learn more about what features you\u2019d like to see. Thanks!","title":"Show HN: Podverse \u2013 AI Superpowers for Podcasters","updated_at":"2024-10-04T19:35:09Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"Jpoliachik"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["deepgram"],"value":"Hi HN, I'm JP

I wanted to automate my voice notes - pocket the phone (or watch), walk and talk, then send the transcript to my agent for processing.\nSo I built Ramble for myself.

It's a simple app. An on-ramp into your systems.\nFree. Open source. No accounts.

Why a new app?\n1) Apple Voice Memos are great, but no way to automate the result. You're also stuck with Apple's transcription (it got slightly better in iOS26 but still not the best)\n2) Voice transcription in Claude / ChatGPT / messaging apps is built for short back-and-forth chat, not capturing a 10-minute walking ramble. They also don't let you record offline on Apple Watch.\n3) I wanted a minimal distraction-free app to record thoughts and trust they'll be delivered. Nothing more.

I've been using Ramble on my phone & watch for months. Talk-to-think became a regular habit and I started getting real value from it. I highly recommend long-form talking to voice notes as a way to process your thoughts, regardless!

Some technical notes:

Native iPhone + Apple Watch, SwiftUI, no third-party deps. App is < 2MB.

Recordings and transcripts stored locally on device. Watch recordings work offline and sync to the phone when in range.

Default transcription is Apple's on-device SFSpeechRecognizer (free, offline). Cloud models (Groq, Deepgram, OpenAI, etc) are optional and route through a stateless Cloudflare Worker that forwards audio and returns text. Nothing persisted server-side. It's all in the repo if you want to verify.

The proxy uses Apple App Attest instead of accounts: each request carries a DCAppAttestService assertion proving it came from a real copy of Ramble on a real device. That's how the cloud tier is gated without ever asking for an email.

Webhooks are sent from device - the transcript POSTs to a URL, signed HMAC-SHA256 with a per-user secret so the receiver can verify it. Pipe it into n8n, an agent, whatever.

Landing page: https://goodloop.dev/ramble/

Repo: https://github.com/Jpoliachik/ramble-ios

App Store: https://apps.apple.com/us/app/ramble-voice-notes/id676185775...

Feedback welcome"},"title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: Ramble, iOS voice notes that send webhooks"}},"_tags":["story","author_Jpoliachik","story_47976481","show_hn"],"author":"Jpoliachik","created_at":"2026-05-01T16:14:49Z","created_at_i":1777652089,"num_comments":0,"objectID":"47976481","points":4,"story_id":47976481,"story_text":"Hi HN, I'm JP

I wanted to automate my voice notes - pocket the phone (or watch), walk and talk, then send the transcript to my agent for processing.\nSo I built Ramble for myself.

It's a simple app. An on-ramp into your systems.\nFree. Open source. No accounts.

Why a new app?\n1) Apple Voice Memos are great, but no way to automate the result. You're also stuck with Apple's transcription (it got slightly better in iOS26 but still not the best)\n2) Voice transcription in Claude / ChatGPT / messaging apps is built for short back-and-forth chat, not capturing a 10-minute walking ramble. They also don't let you record offline on Apple Watch.\n3) I wanted a minimal distraction-free app to record thoughts and trust they'll be delivered. Nothing more.

I've been using Ramble on my phone & watch for months. Talk-to-think became a regular habit and I started getting real value from it. I highly recommend long-form talking to voice notes as a way to process your thoughts, regardless!

Some technical notes:

Native iPhone + Apple Watch, SwiftUI, no third-party deps. App is < 2MB.

Recordings and transcripts stored locally on device. Watch recordings work offline and sync to the phone when in range.

Default transcription is Apple's on-device SFSpeechRecognizer (free, offline). Cloud models (Groq, Deepgram, OpenAI, etc) are optional and route through a stateless Cloudflare Worker that forwards audio and returns text. Nothing persisted server-side. It's all in the repo if you want to verify.

The proxy uses Apple App Attest instead of accounts: each request carries a DCAppAttestService assertion proving it came from a real copy of Ramble on a real device. That's how the cloud tier is gated without ever asking for an email.

Webhooks are sent from device - the transcript POSTs to a URL, signed HMAC-SHA256 with a per-user secret so the receiver can verify it. Pipe it into n8n, an agent, whatever.

Landing page: https://goodloop.dev/ramble/

Repo: https://github.com/Jpoliachik/ramble-ios

App Store: https://apps.apple.com/us/app/ramble-voice-notes/id676185775...

Feedback welcome","title":"Show HN: Ramble, iOS voice notes that send webhooks","updated_at":"2026-05-02T03:37:42Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"stephanieriggs"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["deepgram"],"value":"AI is awesome in the terminal but many people aren't comfortable using command lines - which means they can't access or utilize the full power of AI.

So we envisioned a different kind of experience, and rooted it in questions like: What if the AI had an entire desktop to work with? What if AI was the experience rather than the tool? And what if that experience was fun, delightful, and intuitive? Today we dropped a first look video of what we created.

PearlOS is a browser based desktop environment where the AI companion (Pearl) talks to you to open apps, manage windows, build characters, research, takes notes, searches the web, and controls the whole UI. You just talk to her and things happen on screen.

It's early. But the core works and our entire (small) team uses it for daily tasks\nWalkthrough Video: https://www.youtube.com/watch?v=aKO52ox0dx0\nGitHub: https://github.com/NiaExperience/PearlOS/

We want to build this out into \u201cpersonal pearls\u201d that everyone can have with no coding/AI experience at all needed.

Looking for architecture feedback, contributors, and honest criticism.

What you get when you open it: * A desktop environment with apps (notes, browser, YouTube, file manager, calculator, music player), all * A voice you can talk to naturally. Interruption handling, turn taking, real conversation. * Persistent memory. Pearl remembers your projects, preferences, and past conversations across sessions. She picks up where you left off. * Sub-agent swarms. Pearl can break complex tasks into parallel agent jobs that run in the background while you keep talking.

Stack: * Next.js frontend (the desktop runs in a browser) * Multi-model routing (fast model for chat, heavier model for complex reasoning, configurable) * Pipecat for real-time voice (Deepgram STT, PocketTTS for local TTS) * OpenClaw for agent orchestration * Bring your own API keys, swap in any LLM

What it's not: * Not another ChatGPT wrapper * Not a terminal with an AI bolted on * Not collecting your data (runs on your hardware, no telemetry)

Thanks everyone!\nStephanie & the PearlOS team"},"title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: PearlOS: we gave AI a talking desktop environment instead of a text box"}},"_tags":["story","author_stephanieriggs","story_47441454","show_hn"],"author":"stephanieriggs","created_at":"2026-03-19T15:49:09Z","created_at_i":1773935349,"num_comments":0,"objectID":"47441454","points":4,"story_id":47441454,"story_text":"AI is awesome in the terminal but many people aren't comfortable using command lines - which means they can't access or utilize the full power of AI.

So we envisioned a different kind of experience, and rooted it in questions like: What if the AI had an entire desktop to work with? What if AI was the experience rather than the tool? And what if that experience was fun, delightful, and intuitive? Today we dropped a first look video of what we created.

PearlOS is a browser based desktop environment where the AI companion (Pearl) talks to you to open apps, manage windows, build characters, research, takes notes, searches the web, and controls the whole UI. You just talk to her and things happen on screen.

It's early. But the core works and our entire (small) team uses it for daily tasks\nWalkthrough Video: https://www.youtube.com/watch?v=aKO52ox0dx0\nGitHub: https://github.com/NiaExperience/PearlOS/

We want to build this out into \u201cpersonal pearls\u201d that everyone can have with no coding/AI experience at all needed.

Looking for architecture feedback, contributors, and honest criticism.

What you get when you open it: * A desktop environment with apps (notes, browser, YouTube, file manager, calculator, music player), all * A voice you can talk to naturally. Interruption handling, turn taking, real conversation. * Persistent memory. Pearl remembers your projects, preferences, and past conversations across sessions. She picks up where you left off. * Sub-agent swarms. Pearl can break complex tasks into parallel agent jobs that run in the background while you keep talking.

Stack: * Next.js frontend (the desktop runs in a browser) * Multi-model routing (fast model for chat, heavier model for complex reasoning, configurable) * Pipecat for real-time voice (Deepgram STT, PocketTTS for local TTS) * OpenClaw for agent orchestration * Bring your own API keys, swap in any LLM

What it's not: * Not another ChatGPT wrapper * Not a terminal with an AI bolted on * Not collecting your data (runs on your hardware, no telemetry)

Thanks everyone!\nStephanie & the PearlOS team","title":"Show HN: PearlOS: we gave AI a talking desktop environment instead of a text box","updated_at":"2026-03-19T19:37:42Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"lharries"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["deepgram"],"value":"Luke from ElevenLabs here. We just released Scribe Realtime v2, our new real-time Speech to Text model.

We rebuilt it from the ground up around a new architecture that uses predictive transcription \u2013 it anticipates the next likely word before it\u2019s spoken. That makes it both fast (around 150 ms latency) and highly accurate.

Scribe v2 Realtime outperforms every low latency transcription model across 30 commonly used EU and Asian languages \u2013 93.5% accuracy, compared to Gemini 2.5 Flash (91.4%), GPT-4o MiniTranscribe (90.7%), and Deepgram Nova 3 (85%).

It supports 90+ languages including English, French, German, Italian, Spanish, Portuguese, Hindi, and Japanese. It\u2019s designed for live, agentic applications like AI voice agents, meeting notetakers, and captioning systems.

We also built an open-source real-time transcription UI component at https://ui.elevenlabs.io/blocks to make it easy to integrate voice into any product.

You can use Scribe v2 Realtime via our API or directly within ElevenLabs Agents.

Try it out here: https://elevenlabs.io/realtime-speech-to-text or read the docs: https://elevenlabs.io/docs/capabilities/speech-to-text

We\u2019d love your feedback!"},"title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: Scribe v2 Realtime \u2013 new SOTA real-time speech to text model"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://elevenlabs.io/realtime-speech-to-text"}},"_tags":["story","author_lharries","story_45905100","show_hn"],"author":"lharries","created_at":"2025-11-12T19:33:24Z","created_at_i":1762976004,"num_comments":0,"objectID":"45905100","points":4,"story_id":45905100,"story_text":"Luke from ElevenLabs here. We just released Scribe Realtime v2, our new real-time Speech to Text model.

We rebuilt it from the ground up around a new architecture that uses predictive transcription \u2013 it anticipates the next likely word before it\u2019s spoken. That makes it both fast (around 150 ms latency) and highly accurate.

Scribe v2 Realtime outperforms every low latency transcription model across 30 commonly used EU and Asian languages \u2013 93.5% accuracy, compared to Gemini 2.5 Flash (91.4%), GPT-4o MiniTranscribe (90.7%), and Deepgram Nova 3 (85%).

It supports 90+ languages including English, French, German, Italian, Spanish, Portuguese, Hindi, and Japanese. It\u2019s designed for live, agentic applications like AI voice agents, meeting notetakers, and captioning systems.

We also built an open-source real-time transcription UI component at https://ui.elevenlabs.io/blocks to make it easy to integrate voice into any product.

You can use Scribe v2 Realtime via our API or directly within ElevenLabs Agents.

Try it out here: https://elevenlabs.io/realtime-speech-to-text or read the docs: https://elevenlabs.io/docs/capabilities/speech-to-text

We\u2019d love your feedback!","title":"Show HN: Scribe v2 Realtime \u2013 new SOTA real-time speech to text model","updated_at":"2026-03-05T22:57:40Z","url":"https://elevenlabs.io/realtime-speech-to-text"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"floriankiem"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["deepgram"],"value":"We started having issues getting our notes done after having conversations with many other builders in 2023 Therefore, my friend and I started building Sona Insight, an AI-powered transcription app (iOS).

I know every "indie-hacker" is doing something similar now (which I want to speak about later in this post), but our app has some features (e.g. creation of summary templates, own AI backend built on top of OpenAI's whisper, and a unique architecture) that make it unique. Everything is auto-saved to your account and synced to the cloud (so when we release our web app you can see and edit everything on your desktop).

As we built this for ourselves the goal never was to make a gigantic thing out of it as we worked on it in the evenings. But the app gets used successfully and we have a lot of subscribers compared to the marketing effort we invested till now.

What I want to leave for discussion here, with the context given, is why on earth are that many people building something so similar? When we started posting about it on X, we only saw apps from 2018 that had another approach, however, in October and November of this year it felt like every second post was about an app built similarly. Looking into it they always seemed to only do one API call to OpenAI or Deepgram directly, but some of them didn't even work right."},"title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: Voice to Text, but User Friendly"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://sona.wtf"}},"_tags":["story","author_floriankiem","story_42272408","show_hn"],"author":"floriankiem","children":[42272428],"created_at":"2024-11-29T09:14:07Z","created_at_i":1732871647,"num_comments":2,"objectID":"42272408","points":3,"story_id":42272408,"story_text":"We started having issues getting our notes done after having conversations with many other builders in 2023 Therefore, my friend and I started building Sona Insight, an AI-powered transcription app (iOS).

I know every "indie-hacker" is doing something similar now (which I want to speak about later in this post), but our app has some features (e.g. creation of summary templates, own AI backend built on top of OpenAI's whisper, and a unique architecture) that make it unique. Everything is auto-saved to your account and synced to the cloud (so when we release our web app you can see and edit everything on your desktop).

As we built this for ourselves the goal never was to make a gigantic thing out of it as we worked on it in the evenings. But the app gets used successfully and we have a lot of subscribers compared to the marketing effort we invested till now.

What I want to leave for discussion here, with the context given, is why on earth are that many people building something so similar? When we started posting about it on X, we only saw apps from 2018 that had another approach, however, in October and November of this year it felt like every second post was about an app built similarly. Looking into it they always seemed to only do one API call to OpenAI or Deepgram directly, but some of them didn't even work right.","title":"Show HN: Voice to Text, but User Friendly","updated_at":"2024-11-29T10:16:43Z","url":"https://sona.wtf"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"omars_"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["deepgram"],"value":"Hey HN,

I\u2019ve been working on a journaling app that takes a different approach via conversational AI: instead of writing or typing, you speak to an AI that listens and talks back, helping you reflect on your thoughts and feelings in real-time.

It\u2019s designed for people who struggle with traditional journaling. A live transcript is available to revisit and annotate later.

The web app is free and still early. I have a lot of ideas for features, but instead of building in isolation, I\u2019d rather shape it with real users\u2014figuring out what actually helps rather than just guessing.

On the tech side, I tested a bunch of options and landed on DailyBots for orchestration, Deepgram for Speech-to-Text, Cartesia for Text-to-Speech, and OpenAI\u2019s 4o-mini for the LLM\u2014balancing latency, quality, and cost.

Try it here: https://innerecho.xyz

I\u2019d love to hear your thoughts\u2014what would make something like this truly useful to you?"},"title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: Conversational AI Journaling App"}},"_tags":["story","author_omars_","story_43447217","show_hn"],"author":"omars_","children":[43447232],"created_at":"2025-03-22T17:19:37Z","created_at_i":1742663977,"num_comments":1,"objectID":"43447217","points":3,"story_id":43447217,"story_text":"Hey HN,

I\u2019ve been working on a journaling app that takes a different approach via conversational AI: instead of writing or typing, you speak to an AI that listens and talks back, helping you reflect on your thoughts and feelings in real-time.

It\u2019s designed for people who struggle with traditional journaling. A live transcript is available to revisit and annotate later.

The web app is free and still early. I have a lot of ideas for features, but instead of building in isolation, I\u2019d rather shape it with real users\u2014figuring out what actually helps rather than just guessing.

On the tech side, I tested a bunch of options and landed on DailyBots for orchestration, Deepgram for Speech-to-Text, Cartesia for Text-to-Speech, and OpenAI\u2019s 4o-mini for the LLM\u2014balancing latency, quality, and cost.

Try it here: https://innerecho.xyz

I\u2019d love to hear your thoughts\u2014what would make something like this truly useful to you?","title":"Show HN: Conversational AI Journaling App","updated_at":"2025-03-23T21:49:24Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"mayowa_osibodu"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["deepgram"],"value":"Keyterm prompting is a valuable way to help your STT better recognize unique terms like brand names etc, but for non-English languages/non-standard accents, providers like Deepgram tend to hallucinate keyterms in STT transcripts. So the output transcript contains the given keyterms, even when those keyterms are not present in the input audio.

I'm currently collecting feedback to improve this product. Right now it cuts down keyterm hallucinations by about 60% on in-house test data, so I'm curious to see how it performs in public.

The product is free to use while in beta (Hindi and Indian-accented English are supported). \nWould love to hear how it performs on your data. Feel free to drop a comment if you\u2019re interested in features like additional language support, streaming and self-hosting."},"title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: Keyterm Filtering for Voice AI"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://aditu.tech/keyterm-filtering"}},"_tags":["story","author_mayowa_osibodu","story_48028931","show_hn"],"author":"mayowa_osibodu","created_at":"2026-05-05T21:36:27Z","created_at_i":1778016987,"num_comments":0,"objectID":"48028931","points":3,"story_id":48028931,"story_text":"Keyterm prompting is a valuable way to help your STT better recognize unique terms like brand names etc, but for non-English languages/non-standard accents, providers like Deepgram tend to hallucinate keyterms in STT transcripts. So the output transcript contains the given keyterms, even when those keyterms are not present in the input audio.

I'm currently collecting feedback to improve this product. Right now it cuts down keyterm hallucinations by about 60% on in-house test data, so I'm curious to see how it performs in public.

The product is free to use while in beta (Hindi and Indian-accented English are supported). \nWould love to hear how it performs on your data. Feel free to drop a comment if you\u2019re interested in features like additional language support, streaming and self-hosting.","title":"Show HN: Keyterm Filtering for Voice AI","updated_at":"2026-05-06T13:56:11Z","url":"https://aditu.tech/keyterm-filtering"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"yujonglee"},"story_text":{"matchLevel":"none","matchedWords":[],"value":"Hi, this is Yujong from Char (https://github.com/fastrepl/char).

This CLI is a repackaging of the audio/transcription pipeline that powers our desktop app. We noticed that some people build lightweight CLIs to record conversations and send them to tools like OpenClaw or other agents for transcription and workflow automation, so we wanted to make this more accessible as a standalone tool.

It provides a simple interface to record audio, capture system output, and run transcription across multiple providers (local: WhisperCpp, Cactus; cloud: Deepgram, Soinox, AssemblyAI, and more), using the same abstraction layer we use internally.

We thought this could be a useful building block for both humans and agents.

Curious to hear what others think!

You can find examples and documentation here: https://cli.char.com

Thank you!"},"title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: Composable CLI for your meeting workflow"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://cli.char.com/"}},"_tags":["story","author_yujonglee","story_47602786","show_hn"],"author":"yujonglee","children":[47605766],"created_at":"2026-04-01T16:10:53Z","created_at_i":1775059853,"num_comments":0,"objectID":"47602786","points":3,"story_id":47602786,"story_text":"Hi, this is Yujong from Char (https://github.com/fastrepl/char).

This CLI is a repackaging of the audio/transcription pipeline that powers our desktop app. We noticed that some people build lightweight CLIs to record conversations and send them to tools like OpenClaw or other agents for transcription and workflow automation, so we wanted to make this more accessible as a standalone tool.

It provides a simple interface to record audio, capture system output, and run transcription across multiple providers (local: WhisperCpp, Cactus; cloud: Deepgram, Soinox, AssemblyAI, and more), using the same abstraction layer we use internally.

We thought this could be a useful building block for both humans and agents.

Curious to hear what others think!

You can find examples and documentation here: https://cli.char.com

Thank you!","title":"Show HN: Composable CLI for your meeting workflow","updated_at":"2026-04-01T20:00:02Z","url":"https://cli.char.com/"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"soroucsh"},"story_text":{"matchLevel":"none","matchedWords":[],"value":"Hi HN, I built VectorVid after repeatedly seeing teams hack together Whisper + vector DB to search inside webinars and demos.

The problem: You have 100+ hours of videos. You want to index them for RAG. But the pipeline is messy\u2014transcription, frame sampling, OCR, chunking, embeddings, then plugging into your own vector DB.

VectorVid does one thing: video \u2192 RAG-ready JSON.

Input: Video URL (webinar, lecture, demo)\nOutput: { chunks: [{ start_sec, end_sec, text, scene_description, ocr_text, embedding }] }

How it works:

Transcript + speaker diarization (Whisper/Deepgram)

Frame sampling (1/5s) + OCR (EasyOCR/Claude)

Scene descriptions for visual context

Embeddings included (OpenAI)

The MVP is a live demo\u2014you can search inside the 2007 iPhone keynote and see the exact JSON API output.

Tech: Next.js frontend, async processing, Supabase pgvector, deployed on Vercel.

Use cases I'm seeing:

SaaS teams: "Search our help videos" \u2192 power internal search/chat

EdTech: "Students find specific slides" \u2192 jump straight to diagrams

Sales: "Did the pricing slide appear?" \u2192 automated demo auditing

Early feedback wanted. Try the demo and let me know what you'd build on top of this."},"title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: VectorVid \u2013 Convert Videos to RAG-Ready Chunks (No UI, No DB)"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://www.vector-vid.com/"}},"_tags":["story","author_soroucsh","story_46493116","show_hn"],"author":"soroucsh","created_at":"2026-01-04T22:39:59Z","created_at_i":1767566399,"num_comments":0,"objectID":"46493116","points":3,"story_id":46493116,"story_text":"Hi HN, I built VectorVid after repeatedly seeing teams hack together Whisper + vector DB to search inside webinars and demos.

The problem: You have 100+ hours of videos. You want to index them for RAG. But the pipeline is messy\u2014transcription, frame sampling, OCR, chunking, embeddings, then plugging into your own vector DB.

VectorVid does one thing: video \u2192 RAG-ready JSON.

Input: Video URL (webinar, lecture, demo)\nOutput: { chunks: [{ start_sec, end_sec, text, scene_description, ocr_text, embedding }] }

How it works:

Transcript + speaker diarization (Whisper/Deepgram)

Frame sampling (1/5s) + OCR (EasyOCR/Claude)

Scene descriptions for visual context

Embeddings included (OpenAI)

The MVP is a live demo\u2014you can search inside the 2007 iPhone keynote and see the exact JSON API output.

Tech: Next.js frontend, async processing, Supabase pgvector, deployed on Vercel.

Use cases I'm seeing:

SaaS teams: "Search our help videos" \u2192 power internal search/chat

EdTech: "Students find specific slides" \u2192 jump straight to diagrams

Sales: "Did the pricing slide appear?" \u2192 automated demo auditing

Early feedback wanted. Try the demo and let me know what you'd build on top of this.","title":"Show HN: VectorVid \u2013 Convert Videos to RAG-Ready Chunks (No UI, No DB)","updated_at":"2026-03-05T23:19:36Z","url":"https://www.vector-vid.com/"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"DoubleThing"},"story_text":{"matchLevel":"none","matchedWords":[],"value":"Curious to hear what approaches people are taking, what the bottlenecks are, and whether anyone here is pushing toward the goal of "AI that understands you, the first time."

I've been diving into the gap between benchmark ASR performance and real-world speech. Models like Whisper and Deepgram show impressive >95% accuracy in ideal conditions. But in the wild \u2014 accents, noisy environments, emotional speech, code-switching, overlapping speakers \u2014 accuracy often drops sharply, often to the mid-80s or worse.

This matters because the next wave of AI won't be chatbots; it will be hands-free, real-time systems in contexts like:

- care work (voice logs)\n- crisis communication\n- home healthcare\n- security rounds\n- field operations\n- "I need help" micro-interactions

In these high-stakes contexts, 85% accuracy means critical information gets lost.

What seems missing today:

- Fine-tuning pipelines for noisy, accented speech\n- Reinforcement learning loops (user corrections \u2192 model improvements)\n- Fast per-speaker adaptation\n- Better handling of disfluencies ("uh," "um," repairs)\n- Scaling-law insights applied to ASR models\n- Evaluation metrics that reflect real environments instead of curated datasets

What I'm trying to understand:

- What prevents ASR from reaching reliable >99% accuracy in real-world conditions?\n- Is the bottleneck the model architecture, data quality, or something else?

Would love to hear from anyone who has:

- Worked on Whisper fine-tuning\n- Tackled multilingual or accented ASR\n- Shipped speech systems in noisy environments\n- Developed conversational (not dictation) ASR models\n- Built correction-feedback training loops\n- Deployed ASR in safety-critical or field environments

What worked? What failed? What surprised you?"},"title":{"matchLevel":"none","matchedWords":[],"value":"Why is real-world ASR still ~85% when lab models claim >95%?"}},"_tags":["story","author_DoubleThing","story_46095355","ask_hn"],"author":"DoubleThing","created_at":"2025-11-30T10:02:36Z","created_at_i":1764496956,"num_comments":0,"objectID":"46095355","points":3,"story_id":46095355,"story_text":"Curious to hear what approaches people are taking, what the bottlenecks are, and whether anyone here is pushing toward the goal of "AI that understands you, the first time."

I've been diving into the gap between benchmark ASR performance and real-world speech. Models like Whisper and Deepgram show impressive >95% accuracy in ideal conditions. But in the wild \u2014 accents, noisy environments, emotional speech, code-switching, overlapping speakers \u2014 accuracy often drops sharply, often to the mid-80s or worse.

This matters because the next wave of AI won't be chatbots; it will be hands-free, real-time systems in contexts like:

- care work (voice logs)\n- crisis communication\n- home healthcare\n- security rounds\n- field operations\n- "I need help" micro-interactions

In these high-stakes contexts, 85% accuracy means critical information gets lost.

What seems missing today:

- Fine-tuning pipelines for noisy, accented speech\n- Reinforcement learning loops (user corrections \u2192 model improvements)\n- Fast per-speaker adaptation\n- Better handling of disfluencies ("uh," "um," repairs)\n- Scaling-law insights applied to ASR models\n- Evaluation metrics that reflect real environments instead of curated datasets

What I'm trying to understand:

- What prevents ASR from reaching reliable >99% accuracy in real-world conditions?\n- Is the bottleneck the model architecture, data quality, or something else?

Would love to hear from anyone who has:

- Worked on Whisper fine-tuning\n- Tackled multilingual or accented ASR\n- Shipped speech systems in noisy environments\n- Developed conversational (not dictation) ASR models\n- Built correction-feedback training loops\n- Deployed ASR in safety-critical or field environments

What worked? What failed? What surprised you?","title":"Why is real-world ASR still ~85% when lab models claim >95%?","updated_at":"2026-03-05T23:10:41Z"}],"hitsPerPage":50,"nbHits":143,"nbPages":3,"page":0,"params":"query=DeepGram&tags=story&hitsPerPage=50&advancedSyntax=true&analyticsTags=backend","processingTimeMS":8,"processingTimingsMS":{"_request":{"roundTrip":16},"afterFetch":{"format":{"highlighting":3,"total":4},"merge":{"total":1},"total":1},"fetch":{"query":5,"total":6},"total":8},"query":"DeepGram","serverTimeMS":13}