--- title: Meta's Muse Voice Transcribe: a real-time speech model... description: Meta's AI research lab released Muse Voice Transcribe, a real-time audio perception model that combines transcription, speaker diarization, and endpointing... canonical: https://daily.dev/posts/meta-s-muse-voice-transcribe-a-real-time-speech-model-that-handles-transcription-diarization-and--y3qwttey5 twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Meta's Muse Voice Transcribe: a real-time speech model that handles transcription, diarization, and endpointing in one pass | daily.dev og:description: Meta's AI research lab released Muse Voice Transcribe, a real-time audio perception model that combines transcription, speaker diarization, and endpointing... og:url: https://daily.dev/posts/meta-s-muse-voice-transcribe-a-real-time-speech-model-that-handles-transcription-diarization-and--y3qwttey5 og:image: https://api.daily.dev/og/posts/Y3QwTTEy5.png og:image:alt: Meta's Muse Voice Transcribe: a real-time speech model that handles transcription, diarization, and endpointing in one pass og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Meta's Muse Voice Transcribe: a real-time speech model that handles transcription, diarization, and endpointing in one pass **[Collections](https://daily.dev/sources/collections)** · 2 min read · 0 upvotes · 0 comments ## Summary Meta's AI research lab released Muse Voice Transcribe, a real-time audio perception model that combines transcription, speaker diarization, and endpointing into a single pass rather than stitching together separate models. It processes audio in 80ms chunks, deciding after each chunk whether to emit text or keep listening, achieving a 3.1% final-transcription word error rate with adaptive delay that beats competing streaming speech-to-text models. It's positioned as a perception layer for voice agents and is available via the Meta Model API, Meta AI for Mac, and Muse Code. ## Content Meta's Superintelligence Labs (MSL) has released Muse Voice Transcribe, its first real-time audio perception model. It leads the AA-WER Streaming benchmark with a 3.1% word error rate, ahead of Cartesia Ink-2, ElevenLabs Scribe v2 Realtime, GPT Live Transcribe, and Gemini 3.5 Transcribe Live. ## What makes it different Most speech-to-text systems transcribe audio and leave the harder problems — figuring out who's speaking, when a turn ends, how long to wait before committing to a word — to separate components or the application layer. Muse handles all of that inside a single streaming model. It processes audio in 80ms chunks and decides after each chunk whether to emit text or keep listening. That timing decision is learned through reinforcement learning, which Meta calls "adaptive delay" — the model figures out when it has enough context to commit to a word without waiting unnecessarily. Other capabilities: - Speaker diarization for 20+ speakers - 70+ language support, including mid-conversation language switching - Works on hour-long conversations - Native endpointing (detecting when a speaker's turn is actually over) ## Availability and pricing Muse Voice Transcribe is available through the Meta Model API, Meta AI for Mac, and Muse Code, priced at $3.00 per 1,000 audio-minutes. Unlike Meta's Llama models, the weights won't be open-sourced. ## Questions this post answers ### What is Meta's Muse Voice Transcribe model and how does it differ from a standard speech-to-text API? Muse Voice Transcribe is Meta's first real-time audio perception model, handling transcription, speaker diarization, and endpointing within a single model rather than as separate post-processing stages. It processes audio in 80ms chunks and decides after each chunk whether to emit text or keep listening, achieving a 3.1% final-transcription word error rate with adaptive delay, ahead of competing streaming speech-to-text models. _Teams building voice agents can track new speech-model releases like this on daily.dev._ ### Where can I access Meta's Muse Voice Transcribe model? It is available through the Meta Model API, Meta AI for Mac, and Muse Code. It functions as a perception layer suited for voice agents, since it combines transcription, diarization, and endpointing in one pass instead of requiring multiple stitched-together models. _Developers evaluating voice-agent tooling can follow access changes for models like this on daily.dev._ ## Community take How the wider developer community reacted, aggregated from 3 discussions and 195 comments across x (as of 2026-09-01). **TL;DR:** Reaction is mostly hype and congratulations toward Meta for shipping fast, with genuine interest in the streaming diarization and adaptive-delay approach, tempered by skepticism about closed weights, pricing at scale, real-world noisy audio, and availability outside the US. **Sentiment:** 55% positive · 25% mixed · 20% skeptical **The case for** - Combining streaming transcription, diarization, and turn-end detection into one model impresses commenters who found post-processing diarization janky before. - Several see the adaptive-delay latency mechanism as a clever solution to balancing speed and accuracy. - The quoted pricing ($0.18/hour) is seen as reasonable if quality matches the demo. **The pushback** - Not releasing open weights (unlike Llama) disappoints people who wanted to self-host or build on it. - Some doubt the happy-path demo will hold up with real overlapping speakers and background noise. - Streaming pricing may not map cleanly to keeping a GPU session open for a whole call versus batch processing. - Limited availability (e.g. not in India) frustrates people who want to try it. - One reply argues competitors already solved real-time streaming and diarization months earlier. **By community** - x (positive): Largely congratulatory and excited replies about the tech and shipping speed, mixed with off-topic spam and some pointed concerns about closed weights and real-world robustness. **Hottest debate:** Whether the impressive benchmark demo will actually hold up in messy, overlapping, noisy real-world audio versus a curated happy-path showcase. **Open questions** - How well does it handle heavy code-switching, e.g. Hindi/English or other regional-language mixing mid-sentence? - Does it provide persistent unique speaker embeddings/IDs rather than just per-session speaker indices? - How is streaming priced relative to batch given the GPU has to stay open for the whole call? - Will weights ever be open-sourced given Meta's Llama precedent? **Highlights** > @finkd Incredible performance that blows the existing SOTA at detecting voices out of the water, but it's a shame you're not releasing an open weights model version > — [full\_kelly\_ on x](https://x.com/full_kelly_/status/2094841343726034978) > @finkd Streaming STT with native diarization in one model is nuts. Diarization post-processing is always the janky part. > — [websterweby on x · 1 points](https://x.com/websterweby/status/2094837161954811958) > @Abhishocked @finkd the $0.10/hr ceiling tracks for batch, but streaming is a different cost shape. you're holding the GPU open for the whole call instead of chewing through audio at 30x realtime. curious if meta prices those separately or just eats it to get this into the glasses. > — [NuvaxisAI on x](https://x.com/NuvaxisAI/status/2094865830362149246) > @finkd Does it return unique embedding for the speaker? If not it’s useless as we can’t dereference against known users. Cc @ElevenLabsDevs @DeepgramAI first one to give me unique IDs - not indexes in the thread - takes the money. > — [PietroCasella on x · 1 comments](https://x.com/PietroCasella/status/2094857830197067940) **Source threads** - [x](https://x.com/finkd/status/2094836602681938385) · 0 points · 194 comments - [x](https://x.com/rohanpaul_ai/status/2094843121658216896) · 0 points · 1 comments - [x](https://x.com/giffmana/status/2094846863878848919) · 0 points · 0 comments --- Tags: [#speech-recognition](https://daily.dev/tags/speech-recognition), [#voice-ai](https://daily.dev/tags/voice-ai) [View this post on daily.dev](https://daily.dev/posts/meta-s-muse-voice-transcribe-a-real-time-speech-model-that-handles-transcription-diarization-and--y3qwttey5) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Meta's Muse Voice Transcribe: a real-time speech model that handles transcription, diarization, and endpointing in one pass","url":"https://daily.dev/posts/meta-s-muse-voice-transcribe-a-real-time-speech-model-that-handles-transcription-diarization-and--y3qwttey5","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/meta-s-muse-voice-transcribe-a-real-time-speech-model-that-handles-transcription-diarization-and--y3qwttey5"},"datePublished":"2026-09-01T17:42:04.923Z","dateModified":"2026-09-01T19:27:44.052Z","description":"Meta's AI research lab released Muse Voice Transcribe, a real-time audio perception model that combines transcription, speaker diarization, and endpointing...","image":"https://pbs.twimg.com/media/HRJftxla4AARaf7.jpg","thumbnailUrl":"https://pbs.twimg.com/media/HRJftxla4AARaf7.jpg","isAccessibleForFree":true,"articleSection":"Collections","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Collections","logo":"https://media.daily.dev/image/upload/s--fk_6ycEi--/f_auto,q_auto/v1780996001/logos/collections?_a=BAMAMiWQ0","url":"https://daily.dev/sources/collections"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/meta-s-muse-voice-transcribe-a-real-time-speech-model-that-handles-transcription-diarization-and--y3qwttey5","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"speech-recognition,voice-ai","timeRequired":"PT2M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Collections","item":"https://daily.dev/sources/collections"},{"@type":"ListItem","position":3,"name":"Meta's Muse Voice Transcribe: a real-time speech model that handles transcription, diarization, and endpointing in one pass"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/meta-s-muse-voice-transcribe-a-real-time-speech-model-that-handles-transcription-diarization-and--y3qwttey5#faq","mainEntity":[{"@type":"Question","name":"What is Meta's Muse Voice Transcribe model and how does it differ from a standard speech-to-text API?","acceptedAnswer":{"@type":"Answer","text":"Muse Voice Transcribe is Meta's first real-time audio perception model, handling transcription, speaker diarization, and endpointing within a single model rather than as separate post-processing stages. It processes audio in 80ms chunks and decides after each chunk whether to emit text or keep listening, achieving a 3.1% final-transcription word error rate with adaptive delay, ahead of competing streaming speech-to-text models. Teams building voice agents can track new speech-model releases like this on daily.dev."}},{"@type":"Question","name":"Where can I access Meta's Muse Voice Transcribe model?","acceptedAnswer":{"@type":"Answer","text":"It is available through the Meta Model API, Meta AI for Mac, and Muse Code. It functions as a perception layer suited for voice agents, since it combines transcription, diarization, and endpointing in one pass instead of requiring multiple stitched-together models. Developers evaluating voice-agent tooling can follow access changes for models like this on daily.dev."}}]} ```