Muse Voice Transcribe

Muse Voice Transcribe is a hosted speech recognition model developed by Meta Superintelligence Labs. Meta released it on September 1, 2026 as the lab's first real-time audio perception model. It converts live or recorded speech into text and can add speaker labels or detect utterance boundaries. Meta made it available through the Meta Model API, Meta AI for Mac, and voice input in Muse Code.^[1][2][6]

The model is part of the Muse Spark family, but it has a narrower purpose than a general multimodal assistant. Its public API accepts audio and returns transcript text. It does not synthesize speech or provide a speech-to-speech conversation interface.^[1][3]

Scroll sideways for more →

Model design

Meta describes Muse Voice Transcribe as an autoregressive multimodal model in the Muse Spark family. The system processes audio in 80 millisecond chunks, or 12.5 chunks per second, and represents each chunk as one soft token. At each step, the autoregressive model either emits a text token or emits a special token that asks for the next audio chunk. A separate end-of-audio token tells the model to emit any remaining text after the input stream stops.^[1]

This design lets the model choose how much speech to hear before committing to a word. Meta calls that behavior "adaptive delay." The company trained it with reinforcement learning, multiplying a word error rate reward by a delay reward. The stated objective is to wait longer when more acoustic or linguistic context is useful and to emit easier words sooner.^[1]

The latency and accuracy tradeoff is a general problem in streaming ASR. A recognizer that waits for more future audio can often resolve ambiguities that are difficult at the instant a word begins, but that wait delays captions and voice-agent responses. Published streaming-ASR research evaluates model quality and output latency together for this reason.^[8][13]

Meta added diarization and endpointing to the same output sequence with task-specific control tokens. A start-of-turn token marks a possible speaker transition, while speaker tokens label preceding audio spans. Speech-onset and speech-endpoint tokens mark the beginning and end of an utterance. Meta says it trained these tasks jointly with transcription and added diarization and endpointing rewards to the ASR objective.^[1]

Meta did not publish a parameter count, a full training-data inventory, or model weights with the launch materials. The public model catalog distinguishes Muse Voice Transcribe as a hosted Meta Model API model from Muse Glimmer, which Meta distributes as open weights. A Meta spokesperson separately told The New Stack that the company would not release open weights for Muse Voice Transcribe.^[3][9]

Transcription modes

The Meta Model API exposes three modes. A caller selects the mode when opening a real-time session or sending a file, and the mode cannot change after a WebSocket session starts.^[2]

Scroll sideways for more →

Endpointing and diarization answer different questions. Endpointing estimates when an utterance has finished, which lets another system decide when to respond. Diarization partitions speech by speaker, often summarized as "who spoke when." A diarization label such as A or B is not a person's verified identity. Meta's API documentation says that labels apply only within the current session and should not be assumed to identify the same person in another session.^[2][10]

In diarization mode, a speaker event labels the audio immediately before that event. It does not by itself mean that speech stopped or that the speaker changed. The same label can occur several times, and consecutive spans with the same label can be merged by the client. Meta recommends endpointing mode, rather than diarization mode, for low-latency voice-command use.^[2]

Partial transcripts and timing

The real-time endpoint can send provisional text while speech is still arriving. In the default cumulative mode, each partial transcript replaces the previous hypothesis for the active turn, so text may be revised or removed. A delta option sends new per-chunk text for supported file-streaming configurations. The final signal depends on the selected mode: push-to-talk returns a transcript marked as final, while endpointing and diarization return a speechComplete event keyed by a turn identifier.^[2]

The API may open a later turn before an earlier turn's completed transcript arrives. Meta therefore tells clients to track endpointing and diarization results by turn ID rather than assume strict completion order. A speech-end event marks a detected boundary but is not itself the final transcript. The model may still add punctuation or capitalization before sending the completed text.^[2]

Muse Voice Transcribe provides start and end times for turns, not individual words. Its launch API does not return word-level timestamps, confidence scores, sound-event labels, emotion labels, or automatic transcript reformatting. Audio-progress messages count how much audio the server has processed; they are not word timestamps.^[2][3]

Language coverage and biasing

Meta says it trained Muse Voice Transcribe on more than 70 languages. For the initial release, the company designated 25 languages as evaluated and recommended those languages for use. The developer documentation lists Arabic, Bengali, Dutch, English, French, German, Hebrew, Hindi, Indonesian, Italian, Japanese, Kannada, Korean, Malay, Mandarin Chinese, Marathi, Polish, Portuguese, Spanish, Tagalog, Tamil, Telugu, Thai, Turkish, and Vietnamese.^[1][2]

The model can transcribe code-switched speech in which a speaker changes languages within or between sentences. A request can include a list of expected languages. This language bias is a hint rather than a restriction, and the model can detect a language when the hint is omitted. A separate keyword list can bias recognition toward names, places, acronyms, product terms, or specialist vocabulary. Meta's documentation warns that biasing does not guarantee an exact spelling.^[1][2]

Training coverage and evaluated coverage are not equivalent. The launch materials do not provide per-language error rates for all 70-plus training languages, and the independent streaming benchmark used for the launch ranking evaluates English speech. The 25-language API list establishes supported launch languages, not equal measured accuracy across them.^[2][8][9]

API endpoints and audio formats

Both public endpoints run the same model but use different transports and audio constraints.^[2]

Scroll sideways for more →

For live audio, authentication is sent in the first WebSocket handshake frame rather than an HTTP authorization header. The client then sends binary PCM frames at approximately real-time speed. Meta's documentation warns against sending buffered audio substantially faster than it can be processed. For a file request, the caller uploads a supported WAV file in a multipart HTTP request and can ask for JSON, plain text, or server-sent events.^[2]

The launch blog says the underlying model natively supports audio longer than one hour and more than 20 speakers without a separate post-processing stage. It includes a 1:00:52 demonstration with 11 listed speakers. Those model-level demonstrations should not be confused with the public endpoint caps: the developer guide limits a single real-time API session to 60 minutes and a file request to 10 minutes.^[1][2]

Evaluation

Meta's launch page said Muse Voice Transcribe ranked first on Artificial Analysis for streaming speech-to-text on September 1, 2026. A live Artificial Analysis leaderboard inspected after launch independently listed a 3.1% final-transcript word error rate, a 0.16 second time to final transcript after detected speech end, and a normalized price of $3 per 1,000 audio minutes. On that snapshot, the 3.1% result was the lowest final-transcript WER among the displayed streaming systems, though several systems returned final text faster.^[7]

Artificial Analysis calculates its streaming index from about eight hours of audio. The composite weights its private AA-AgentTalk set at 50%, an English subset of VoxPopuli at 25%, and Earnings22 corporate calls at 25%. It streams audio in real time, includes network delay, and measures final-transcript latency from the end of speech detected by Silero VAD. Its method normalizes formatting and computes WER as substitutions plus insertions plus deletions divided by the number of words in a verified reference transcript.^[8]

Scroll sideways for more →

The AMI corpus contains recorded meetings with manual speaker-specific transcripts, while VoxConverse contains conversational videos collected under less controlled conditions, including overlapping speech and background noise.^[11][12] A result averaged across those benchmarks is not a universal error rate for every microphone layout or conversation type. Diarization error and transcription WER also measure different failures, so the percentages should not be compared as if they were the same metric.^[8][10]

Benchmark positions can change when new systems or configurations are added. The launch ranking does not establish error rates for the other 24 evaluated API languages, long sessions with more than 20 speakers, or every acoustic environment. Meta's own browser demonstration also warns that generated transcripts may be inaccurate.^[1][7][8]

Pricing and data handling

Meta launched the service at $0.18 per hour of audio processed, equivalent to $3 per 1,000 minutes. Streaming and file transcription have the same rate. The developer guide says billing uses processed audio rounded down to whole seconds, and failed requests that produce no transcript are not billed. Platform free-tier credits can apply.^[2][4]

The API offers a Zero Data Retention setting at the same price as standard processing. Its real-time API reference describes zdrOverride: true as forcing metadata-only logging, while false permits content retention. If the field is unset, the authenticated caller's policy applies. Zero Data Retention therefore concerns content retention and does not mean that the service keeps no operational metadata.^[4][5][6]

The microphone demonstration on Meta's launch page separately says that audio used in that web demo is processed to produce a transcript and is not stored. That statement applies to the demonstration interface and should not be substituted for the data policy of a developer's own API account or application.^[1][5]

Applications and limits

Meta integrated the model with dictation in Meta AI for Mac and voice input in Muse Code at launch. The Mac interface can send dictated text to the active application when the user holds the Function key. The developer API is intended for live captions, dictation, voice-agent input, meeting transcription, and speaker-attributed call transcripts.^[1][2][6]

The public service remains a transcription component rather than a complete voice agent. An application must supply any downstream reasoning, response generation, speech synthesis, storage, or identity management. Speaker labels do not authenticate a person, endpointing does not guarantee that a pause represents a finished thought, and a low aggregate WER does not guarantee exact wording in a particular recording.^{[2][3][8][10]}

At launch, the main documented constraints were the hosted-only distribution, the 60-minute real-time session cap, the 10-minute file cap, support for WAV rather than compressed file uploads, turn-level rather than word-level timing, and the absence of confidence, sound-event, and emotion outputs.^[2][3][9] The lack of a disclosed parameter count or full training-corpus description also limits external analysis of the model beyond its API behavior and the measurements published at release.^[1][3]

References

1. ^Meta AI Research. "Introducing Muse Voice Transcribe." September 1, 2026. research.meta.ai/...roducing-muse-voice-transcribe
2. ^Meta Model API. "Speech to text." Accessed September 2, 2026. dev.meta.ai/...speech-to-text
3. ^Meta Model API. "Models." Accessed September 2, 2026. dev.meta.ai/...models
4. ^Meta Model API. "Pricing and rate limits." Accessed September 2, 2026. dev.meta.ai/...pricing-rate-limits
5. ^Meta Model API. "Realtime transcription API reference." Accessed September 2, 2026. dev.meta.ai/...realtime
6. ^Mark Zuckerberg. "Muse Voice Transcribe is MSL's first real-time audio perception model." X thread, September 1, 2026. x.com/...2094836602681938385
7. ^Artificial Analysis. "Speech to Text AI Model and Provider Leaderboard: Streaming." Accessed September 2, 2026. artificialanalysis.ai/...streaming
8. ^Artificial Analysis. "Speech to Text Benchmarking Methodology." Accessed September 2, 2026. artificialanalysis.ai/...speech-to-text
9. ^Frederic Lardinois. "Meta just beat OpenAI and Google at real-time transcription." The New Stack, September 1, 2026. thenewstack.io/meta-muse-voice-transcribe
10. ^Tae Jin Park, Naoyuki Kanda, Dimitrios Dimitriadis, Kyu J. Han, Shinji Watanabe, and Shrikanth Narayanan. "A Review of Speaker Diarization: Recent Advances with Deep Learning." Computer Speech & Language 72, 2022. doi.org/...j.csl.2021.101317
11. ^Jean Carletta et al. "The AMI Meeting Corpus: A Pre-announcement." Machine Learning for Multimodal Interaction, 2006. doi.org/...11677482_3
12. ^Joon Son Chung, Jaesung Huh, Arsha Nagrani, Triantafyllos Afouras, and Andrew Zisserman. "Spot the Conversation: Speaker Diarisation in the Wild." Interspeech 2020. doi.org/...Interspeech.2020-2337
13. ^Haibin Miao et al. "Low-latency transformer model for streaming automatic speech recognition." Electronics Letters 58, 2022. doi.org/...ell2.12349