Meta Superintelligence Labs Releases Muse Voice Transcribe: One Real-Time Model for Streaming ASR, Diarization, and Endpointing
Most manufacturing voice stacks are three methods stitched collectively. One mannequin transcribes, a second separates audio system, and a detector decides when the person stopped speaking. Each hand-off provides latency and a brand new failure mode.
Muse Voice Transcribe, introduced by Meta Superintelligence Labs this week, collapses these three jobs right into a single autoregressive mannequin. Meta calls it its first real-time audio notion mannequin. It performs streaming ASR, speaker diarization for 20+ audio system, and endpointing in a single cross, with no required post-processing.
Is it deployable? Yes, however solely as a hosted API. It is stay on the Meta Model API as muse-voice-transcribe-1.0 at $3.00 per 1,000 audio minutes ($0.18 per hour), and it already powers dictation in Meta AI for Mac and Muse Code. No weights have been launched, so there isn’t any self-hosted path.
Streaming ASR as the muse
Muse Voice Transcribe is an autoregressive multimodal mannequin from the Muse Spark household. Audio arrives in 80ms chunks at 12.5 Hz. Each chunk is reworked right into a single comfortable token.
After each chunk the mannequin makes one binary alternative. It both predicts a <|next_audio|> token and retains listening, or it emits a textual content token. When the mannequin predicts <|next_audio|>, that token is changed by the precise subsequent audio chunk within the enter. When the stream ends, an <|empty_audio|> token is inserted, and the mannequin flushes all remaining textual content with out requesting extra audio.
Listening and writing share one decoder loop, so there isn’t any separate alignment stage to float.
Adaptive delay, educated with RL
Because the mannequin controls when it listens, it additionally controls how a lot audio context sits behind every phrase. Meta calls that hole ‘delay.’ Longer delay means a extra correct transcript and larger latency.
Instead of fixing that trade-off, Meta trains it. Reinforcement studying combines a phrase error fee reward and a delay reward multiplicatively, producing a coverage that varies delay per phrase by issue. Meta reviews this places the mannequin on the Pareto entrance for pace in opposition to accuracy, measured by time to closing transcription, forward of the earlier frontier shaped by Soniox, Cartesia, and ElevenLabs methods.
Diarization and endpointing are extra tokens
Meta didn’t add a second mannequin for speaker attribution. It added particular tokens to the identical stream.
For diarization, a <|start_of_turn|> token marks a possible speaker change, and a <|speaker_{A-Z}|> tag identifies the speaker. The flip token fires as quickly as a change is feasible, whereas the speaker tag is delayed to the tip of the chunk. Audio from one speaker might be cut up throughout a number of segments that each one resolve to the identical tag.
For endpointing, <|speech_onset|> marks the beginning of speech and <|speech_endpoint|> marks the purpose the place the person completed. Both duties are educated collectively with streaming ASR, utilizing further rewards layered on prime of the ASR reward.
Capabilities
The mannequin was educated on 70+ languages, of which 25 are extensively verified and really useful at launch. Code-switching is native, each inside a sentence and between sentences, which issues for bilingual audio system who combine languages mid-clause. Accuracy might be improved additional with language, key phrase, and context biasing.
Long-context dealing with is a sensible differentiator. Meta states the mannequin natively helps audio enter exceeding one hour and 20+ audio system, with no required post-processing step.
Benchmarks
Meta reviews first place on Artificial Analysis for streaming speech-to-text and on public diarization benchmarks, as of September 1, 2026.
On Artificial Analysis AA-WER Streaming, Muse Voice Transcribe information 3.1% final-transcript WER at 0.16s after finish of speech. Cartesia Ink-2 with semantic endpoints is 3.4% at 0.43s. ElevenLabs Scribe v2 Realtime is 3.6% at 0.14s. Cartesia Ink-2 with exterior endpoints is quickest at 0.07s however least correct at 4.0%. On first partial transcript, Muse Voice Transcribe information 3.6% WER at 0.13s.
On diarization, Meta reviews a 17.5% common diarization error fee throughout AMI-IHM, AMI-SDM, and VoxConverse. Five different methods in the identical chart vary from 21.1% to twenty-eight.6%.
Price is the opposite axis. At $3.00 per 1,000 minutes, it undercuts Cartesia Ink-2 at $4.00 and is lower than half the $6.50 for ElevenLabs Scribe v2 Realtime and Deepgram Flux.