Muse Voice Transcribe Is Live: Meta's First Real-Time Audio Perception Model

Muse Voice Transcribe, the first real-time audio perception model from Meta Superintelligence Labs, shipped on September 1, 2026, and it is priced like a spoiler: $3.00 per 1,000 minutes of audio — about $0.18 an hour — on the Meta Model API. In one streaming pass it does what most transcription stacks split across three systems: automatic speech recognition, speaker diarization across more than 20 voices, and endpointing, the job of deciding when a speaker has actually finished rather than paused mid-thought. Meta says the model tops the Artificial Analysis streaming speech-to-text leaderboard at a 3.1% word error rate on final transcripts, delivered 0.16 seconds after the speaker stops talking.

That last number needs its honest label before it does any work: it is Meta's launch-day claim, pointing at an independent leaderboard, for a model that had been callable for less than a day when the announcement went out. Nobody outside the lab has reproduced the 3.1% yet, and the accuracy claim is one thing while the rest of the pitch — 70+ trained languages, native code-switching, reinforcement-learned "adaptive delay" — is a separate set of vendor statements in the same boat. Everything in this post is the state of the model on day one, with vendor numbers labeled as vendor numbers.

The numbers that matter on day one

• Price — $3.00 per 1,000 audio-minutes (about $0.18 per audio-hour) on the Meta Model API, undercutting every major streaming transcription API we track.

• Accuracy claim — 3.1% WER on final transcripts at 0.16 seconds after end of speech; 3.6% WER on first partial transcripts at 0.13 seconds. Vendor-reported, citing the Artificial Analysis streaming leaderboard.

• Diarization — 20+ speakers, emitted streaming with no separate post-processing step (Meta reports an average diarization error rate of 17.5%).

• Languages — trained on 70+; 25 "extensively verified" at launch; seamless code-switching within and between sentences.

• Context — handles audio longer than one hour; language, keyword, and context biasing for jargon-heavy domains.

What "audio perception model" means here

The architecture is the story. Muse Voice Transcribe consumes audio in 80-millisecond chunks and streams words out as they stabilize, which is what "real-time" means in practice. The interesting part is how it decides when to commit a word. Rather than a fixed latency budget — the standard trade where a recognizer either commits early and guesses or waits longer and hedges — the model uses what Meta calls "adaptive delay": it moves quickly through easy speech and holds onto more audio context for words it is less sure about. The delay policy itself was learned through reinforcement learning, so the model is effectively balancing speed and accuracy per-word instead of applying one global setting.

That distinction is why Meta calls Muse Voice Transcribe an "audio perception model" rather than an ASR model. The output stream is a single live transcript that also carries who is speaking and when the utterance is complete — three labels that streaming systems usually assemble by gluing a recognizer to a voice-activity detector and a diarization model, each adding its own latency and its own failure modes.

Diarization and endpointing, built in

Speaker diarization is the part most worth scrutinizing, because it is the feature where streaming products most often overstate. Meta says the model separates 20+ speakers in a single recording and reports an average diarization error rate of 17.5%, which it contrasts with a 21.1–28.6% range it attributes to competitor systems. Both numbers are vendor-published on launch day, and no independent evaluation has confirmed the 17.5% yet. What is structurally real is that the diarization runs inside the same streaming pass as the recognition — there is no "upload the audio, wait, get speakers back" second step, which is the design choice that makes 20+ speaker tracking plausible in a live setting at all.

Endpointing is the quieter capability and arguably the most useful. Transcription APIs that expose raw streaming ASR force the caller to implement end-of-utterance detection themselves, which is why voice agents so often cut people off or leave dead air. Muse Voice Transcribe decides when a turn is complete and emits that boundary as part of the stream, so an application gets a clean "this speaker is done" signal without building a VAD layer.

The multilingual claim, read carefully

Meta trained the model on 70+ languages but validates 25 at launch. Those are different things: "trained on" means the data included them; "extensively verified" is the subset Meta actually stands behind with internal testing. For production use, the 25-verified list is the real coverage, and the gap between 70 and 25 is worth knowing before you route 40 languages through it. What is genuinely unusual is the code-switching story — the model is designed to switch languages mid-sentence and mid-utterance without being told, which is a real pain point in multilingual regions and something most single-language ASR products simply cannot do. Language, keyword, and context biasing round out the customization story: you can bias recognition toward a domain vocabulary, which is the feature that makes streaming ASR usable in medical, legal, and support queues.

Three surfaces, one model

Muse Voice Transcribe ships on three surfaces at once. The paid one is the Meta Model API at $3.00 per 1,000 audio-minutes. The consumer one is Meta AI for Mac, where holding the Fn key dictates into any application — Meta's answer to Apple's built-in dictation, and the most visible real-world deployment of the model on day one. The developer one is Muse Code, Meta's coding agent, where voice commands drive multi-agent sessions and refactoring flows. One model powering a dictation feature and a coding agent is the productized version of the "one streaming model, many surfaces" pitch.

What the price undercuts

At $0.18 per audio-hour, Muse Voice Transcribe undercuts the streaming transcription field on list price. The comparison set as we track it: Google's Gemini 3.5 Transcribe Live is about $9 per 1,000 minutes, ElevenLabs' Scribe v2 Realtime is listed at $6.50 per 1,000 minutes, Cartesia's Ink-2 at $4.00, and OpenAI's GPT Live Transcribe at $17.00. Those are the vendors' published numbers, and several of those models have independent accuracy evidence that Muse does not have yet — price leadership on day one is not the same as a settled leaderboard. But a streaming model with built-in diarization and endpointing entering at roughly a fifth to a tenth of the price of the incumbent streaming APIs is the kind of number that resets expectations regardless.

The honest caveats

Muse Voice Transcribe is proprietary, and the weights are not published — the "run it yourself" option does not exist, and there is no open-weights path for audit or fine-tuning. The accuracy claim is launch-day and unreproduced. The 25 verified languages may not match the 70+ "trained on" figure for low-resource coverage. And the diarization benchmark, however promising, is a vendor measurement until an independent run confirms it. For teams whose only requirement is accurate batch transcription of pre-recorded audio, cheaper and better-audited options still exist — the streaming pitch is about real-time interaction, not about beating the batch transcription world on cost per audio-hour.

What to watch next

Four things will decide whether this launch is a moment or a trend. First, whether the Artificial Analysis streaming leaderboard actually lists Muse Voice Transcribe at #1 after independent verification — the launch-day claim needs a settled board entry. Second, whether the 20+ speaker diarization survives contact with real, noisy meetings. Third, whether Meta's Mac dictation surface converts into developer adoption of the API. Fourth, how the incumbents respond on price, because a streaming model with diarization and endpointing at $0.18 an hour puts visible pressure on everyone above it. When Meta opens the model to additional serving partners, a pass-through gateway such as OrcaRouter — which forwards provider list prices at 0% markup — would surface the same $3.00 per 1,000 minutes figure with no reseller padding, the day it lands. Until then, the only place to call Muse Voice Transcribe is Meta's own API.