Meta Superintelligence Labs released Muse Voice Transcribe on September 1, 2026, its first real-time audio perception model, and says it ranks first on Artificial Analysis for streaming speech-to-text. The company will not publish the model’s weights.
The Big Picture
- Meta Superintelligence Labs launched Muse Voice Transcribe, a streaming model that transcribes speech, separates speakers, and detects when a talker stops.
- Muse Voice Transcribe ranks first on the Artificial Analysis streaming leaderboard, ahead of models from Cartesia, ElevenLabs, OpenAI, and Google.
- Meta trained the model on more than 70 languages but validated 25 of them at launch.
- Meta AI for Mac now dictates into any application when a user holds the Fn key.
- Meta will not release open weights for this model, breaking from its Muse Glimmer family.
Meta takes the top spot on streaming transcription accuracy
Artificial Analysis scores streaming transcription with a test it calls AA-WER Streaming. Muse Voice Transcribe records a 3.1% word error rate there, ahead of Cartesia Ink-2 at 3.4%, ElevenLabs Scribe v2 Realtime at 3.6%, GPT Live Transcribe at 3.9%, and Gemini 3.5 Transcribe Live at 4%.
That test covers English speech only. The 70 language training figure sits outside what the leaderboard measures, so the accuracy lead is an English result and carries no claim about the other languages.
Speaker separation is the weaker column. Meta leads public diarization benchmarks at a 17.5% error rate, a mark that leaves speaker attribution well short of solved across the whole field.
Muse Voice Transcribe is MSL's first real-time audio perception model — rolling out today. SOTA in streaming speech-to-text, it handles speaker diarization, and endpointing natively in a single model. pic.twitter.com/LViMDSkbim
— Mark Zuckerberg (@finkd) September 1, 2026
Adaptive delay decides how long the model listens
The model reads audio in 80ms chunks at 12.5 Hz and turns each chunk into one soft token. At every chunk it either emits a text token or predicts a next_audio token and keeps listening, so the model itself controls how much context it hears before committing a word.
That choice is the accuracy tradeoff. “The longer the model waits to predict, the more accurate the transcript, but the higher the latency,” Meta said in its research announcement.
Meta trained the behavior with reinforcement learning, combining a word error rate reward and a delay reward multiplicatively so the model stretches its wait only on hard words. Speaker labeling and end-of-speech detection ride on the same scheme, using start_of_turn and speaker tags for the first, speech_onset and speech_endpoint for the second.
Closed weights break the Muse Glimmer pattern
Meta will not release the weights for this model, a spokesperson told The New Stack, which separates it from the Muse Glimmer line that shipped openly. Developers reach it through the Meta Model API at $3 per 1,000 audio-minutes, equal to $0.18 per hour, and it already runs inside Meta AI for Mac and Muse Code.
That gap matters for anyone who cannot move audio off their own hardware. Meta’s broader AI adoption numbers have climbed on consumer surfaces, and this release stays on the same hosted path. Rivals have made the same call, including Microsoft’s MAI-Voice-1 launch.
Teams already routing audio through a transcription vendor have two checks worth running today. Confirm the languages you handle sit inside the validated set and not the wider training list, and confirm your data policy permits streaming raw audio to a hosted endpoint, because no local weights option exists here.
Four questions stay open:
- Which languages fall outside the validated set, and how far behind are they?
- Which datasets produced the 17.5% diarization figure?
- Does the audio-minute price cover speaker labeling and endpointing, or transcription alone?
- Will Meta offer a self-hosted path for buyers in regulated sectors?
SQ Magazine’s Takeaway
The accuracy lead is real and narrow. Meta cleared the next model by 0.3 percentage points on an English-only test, and the stronger draw is the bundle: transcription, speaker labels, and end-of-speech detection from a single model with no post-processing stage. Voice agent teams currently stitch those three pieces together themselves, and that stitching job is what this release removes.
The weights decision is the part worth watching. Meta built much of its developer goodwill on open releases, and holding this one back converts a research result into a metered API line item.
What’s next: the validated language list should grow past its launch size, competitors will publish streaming numbers measured against Meta’s new mark, and procurement teams in regulated sectors will press for a self-hosted option before they sign.
