What Meta shipped
Muse Voice Transcribe is the first "real-time audio perception model" out of Meta Superintelligence Labs, and an autoregressive multimodal model from the Muse Spark family rather than a bolt-on speech engine. Three jobs happen inside one pass, with no separate post-processing stage:
- Streaming recognition — text as words are said, not after the file ends
- Speaker diarization — separating recordings with more than 20 voices
- Endpointing — deciding when someone has actually stopped, rather than waiting out a silence timer
Trained across more than 70 languages with 25 "extensively verified" at launch, it handles audio longer than an hour and switches languages inside a sentence without being told to. It is live today in the Meta Model API, in Meta AI for Mac, and in Muse Code.
How adaptive delay actually works
The mechanism is worth understanding because it is the whole reason the model sits where it does on the chart. Audio arrives in 80-millisecond chunks, twelve and a half per second, each compressed into a single soft token. At every chunk the model makes one choice: emit a text token, or emit a placeholder that asks for the next chunk of audio. When the audio stops, an "empty audio" token tells it nothing more is coming and it flushes whatever text it is still holding.
Because the model decides how much audio to hear before committing to a word, it controls its own latency per word. Easy words go out almost immediately; hard words get more context. Meta calls this adaptive delay, and the balance is not hand-tuned — it is learned during reinforcement learning, where the accuracy reward and the delay reward are multiplied rather than added, so the model cannot win by sacrificing one for the other. Speaker detection uses a similar mechanism.
The practical consequence: Muse is not the fastest model on the board and not trying to be. It is trying to be the most accurate model that is still fast enough, and the numbers say it got there.
The streaming leaderboard, in full
Artificial Analysis publishes a streaming board separately from its batch one. Here are all eight models with comparable public figures as of 1 September 2026, including the two columns most launch coverage drops.
| Model | AA-WER Streaming | Latency to final | Per 1,000 min |
|---|---|---|---|
| Muse Voice Transcribe (Meta) | 3.1% | 0.16 s | $3.00 |
| Cartesia Ink-2 (semantic endpoints) | 3.4% | 0.43 s | $4.00 |
| ElevenLabs Scribe v2 Realtime | 3.6% | 0.14 s | $6.50 |
| GPT Live Transcribe (OpenAI) | 3.9% | — | $17.00 |
| Gemini 3.5 Transcribe Live (Google) | 4.0% | 0.40 s | $9.00 |
| Cartesia Ink-2 (external endpoints) | 4.0% | 0.07 s | $4.00 |
| Inworld STT 1 Realtime | 4.2% | 0.07 s | $1.40 |
| Deepgram Flux | 7.4% | 0.02 s | $6.50 |
Muse is first on accuracy and second-cheapest. That is not a marketing frame; it is where the point sits on the chart, near the corner you want. OpenAI's entry is the surprise in the other direction: fourth on accuracy at more than five times Muse's price.
One methodological detail that matters: Artificial Analysis reports two streaming error rates — at the final transcript after speech ends, and at the first partial. Every figure here is the final-transcript number. Partial-transcript error is higher for every model, because partials are by definition uncommitted.
Three caveats on "number one"
The number nobody is leading with
| Batch (AA-WER v2) | Streaming (AA-WER Streaming) | ||
|---|---|---|---|
| Fun-Realtime-ASR-preview | 1.7% | Muse Voice Transcribe | 3.1% |
| Scribe v2 (ElevenLabs) | 2.2% | Cartesia Ink-2 (semantic) | 3.4% |
| MAI-Transcribe-1.5 (Microsoft) | 2.4% | ElevenLabs Scribe v2 Realtime | 3.6% |
| Smallest AI Pulse Pro | 2.4% | GPT Live Transcribe | 3.9% |
| Gemini 3.5 Transcribe | 2.6% | Gemini 3.5 Transcribe Live | 4.0% |
| Whisper Large v3 | ~4% | Deepgram Flux | 7.4% |
The best real-time model on the planet, launched today, is less accurate than five models that wait for the file to finish. "Meta beat OpenAI and Google" is true within streaming. Read across the boards and the sentence becomes: if you do not need words appearing while someone is still speaking, reaching for a streaming model is choosing a worse transcript for nothing.
Google's own pair makes the cost concrete. Gemini 3.5 Transcribe scores 2.6% in batch; its Live variant scores 4.0% in streaming. Same vendor, same month, 1.4 points apart — and the Live version costs nearly twice as much per minute.
This is the same trap we wrote about when Gemini 3.5 Transcribe launched, from the other side: comparisons that put a score from one benchmark next to a score from another and let the reader assume they share a scale. Here the scales genuinely match, which is exactly what makes the gap worth stating out loud.
What each model actually is
A row on a leaderboard tells you a number. It does not tell you what the thing is built for, and that is usually what decides the choice.
Muse Voice Transcribe — accuracy first, everything in one model
The one to beat on accuracy, and the only model here that does recognition, 20+ speaker diarization and endpointing in a single pass. 0.16 seconds to a final transcript is mid-pack. $3 per 1,000 minutes is second-cheapest. English-only benchmark, closed weights, and a diarization error rate you have to plan around. Best when the transcript will be read and the audio is English.
Cartesia Ink-2 — built for agents, with two personalities
Cartesia positions Ink-2 for voice agents, and it appears on the board twice because it can be run two ways. With semantic endpoints, the model itself signals when a turn ends — Cartesia's phrase is that it determines turn end "by meaning, not silence," so a pause mid-thought does not make an agent jump in. That earns 3.4% accuracy at 0.43 seconds. With external endpoints, a separate voice-activity detector calls the turn and the model just transcribes: 4.0% at 0.07 seconds. Same weights, different job. Cartesia claims 88 ms turn detection on its own page; it publishes no per-minute price, so the $4 figure is Artificial Analysis's normalisation of plan-based pricing.
ElevenLabs Scribe v2 Realtime — the balanced one
3.6% at 0.14 seconds, which is the best latency of any model under 4% error. ElevenLabs claims under 150 ms on its own page, and 90-plus languages. $0.40 per hour list price works out to about $6.67 per 1,000 minutes, close to AA's $6.50. Its batch sibling, Scribe v2, sits at 2.2% on the other board — a 1.4-point premium for streaming, the same size as Google's.
GPT Live Transcribe — OpenAI's entry, and the expensive one
3.9% on the streaming board at $17 per 1,000 minutes — $0.017 a minute on OpenAI's pricing page — which is more than five times Muse and nearly twice Gemini Live. Artificial Analysis does not surface a final-transcript latency figure for it in its highlights, so that cell is blank rather than guessed. If OpenAI is already your stack, the integration cost may outweigh the per-minute one; on the numbers alone it is hard to pick.
Gemini 3.5 Transcribe Live — Google's streaming variant
4.0% at 0.40 seconds for around $9 per 1,000 minutes — Google prices audio in and text out separately, and the blended rate on its own pricing page is roughly $0.009 per minute. Slower and less accurate than Muse and three times the price. Its batch sibling is the model to want from Google; the Live variant exists for when you need streaming from the same account.
Inworld STT 1 Realtime — the cheap fast one
4.2% at 0.07 seconds for $1.40 per 1,000 minutes on AA's normalisation — Inworld's own list is $0.10 per hour on paid plans, $0.15 on demand. Cheapest on the board by a wide margin, and near the latency floor. Inworld says 30 languages. The trade is a full point of accuracy against Muse; for a voice agent where the LLM downstream can absorb some errors, that may be fine.
Deepgram Flux — the latency floor
0.02 seconds to a final transcript, eight times faster than Muse, at 7.4% error — more than double. Deepgram's own pricing lists Flux English at $0.0065 per minute streaming, matching AA's $6.50. This is a model built for one thing: an agent that must decide, right now, whether to start talking. It is not what you want for a transcript anyone will read. Deepgram's readable-transcript model is Nova-3, and that one lives on the batch comparison.
Latency is an axis, not a footnote
Put the two ends of the board next to each other. Deepgram Flux: 0.02 seconds, 7.4% wrong. Muse: 0.16 seconds, 3.1% wrong. Eight times slower, less than half the errors. Neither is the right answer in general.
A voice agent deciding when to interrupt lives on the 0.02. A meeting transcript someone reads next week does not care about the difference between 0.02 and 0.16 and cares a great deal about the error rate. The board is not a ranking; it is a frontier, and where you land on it is a product decision rather than a benchmark result. That is also why Cartesia ships two modes of one model.
Meta shipped a Mac dictation app too
The part with the most direct consequences for anyone reading this on a Mac: Muse now powers dictation inside Meta AI for Mac, the desktop app Meta released on 19 August as a free beta. Hold a key, speak, and the text lands in whatever application has focus — Mail, a document, a code editor.
That is the same interaction MetaWhisp has, so it is worth being precise about what differs rather than waving it away:
| Meta AI for Mac | MetaWhisp | |
|---|---|---|
| Where audio goes | Meta's servers | Stays on your Mac by default |
| Model | Muse Voice Transcribe, 3.1% streaming | Whisper large-v3-turbo locally, ~4% batch |
| Weights | Closed | Open — Whisper weights are public |
| Account | Meta account | None |
| Works offline | No | Yes |
| Requirements | Apple silicon, macOS 15+ | Apple silicon, macOS 14+ |
| Price | Free beta | Free |
Meta's model is more accurate than the one we run on your machine. That is the honest read and there is no way around it: a 3.1% streaming model beats a ~4% local one. What you trade for those 0.9 points is that every word you dictate goes to Meta, and nothing works on a plane.
Which of those you care about is genuinely yours to decide. If your dictation is Slack messages and shopping lists, the accuracy is worth more than the privacy. If it is client notes, medical language or anything under an NDA, it is not close.
Closed weights, and why that is the part that lasts
Benchmarks move every month; this decision does not. And it matters more for a speech model than for a text one. Text you send to an API is text you chose to write. Audio carries the other people in the room, the background conversation, the thing said before the recording was supposed to start. A closed, cloud-only speech model means all of that is somebody else's now, and the decision is one-way.
Open weights would have made Muse a genuinely different proposition — someone would have had it on Apple silicon within a fortnight, the way Whisper was ported. That did not happen here, and Meta chose for it not to.
Which one, for which job
Frequently asked questions
How accurate is Muse Voice Transcribe?
3.1% word error rate on the AA-WER Streaming index, first place as of 1 September 2026. Behind it: Cartesia Ink-2 semantic endpoints at 3.4%, ElevenLabs Scribe v2 Realtime at 3.6%, GPT Live Transcribe at 3.9%, Gemini 3.5 Transcribe Live at 4.0%. English only.
Is it more accurate than Whisper?
On the numbers, yes — 3.1% streaming against roughly 4% for Whisper Large v3 on the batch board, and both boards use the same three datasets at the same weights. A streaming model beating a batch one is unusual. It is still behind five batch models at 1.7-2.6%.
What does it cost?
$3 per 1,000 audio-minutes via the Meta Model API — $0.003 a minute, about $0.18 an hour. Only Inworld is cheaper at $1.40. Gemini Live is $9; OpenAI's GPT Live Transcribe is $17.
Can I run it locally?
No. A Meta spokesperson confirmed the weights will not be released. Muse Glimmer was released openly under Apache 2.0; this model was not. It runs only through the Meta Model API, Meta AI for Mac and Muse Code.
How good is the speaker diarization?
20+ speakers, natively, in the same pass. Meta says it leads real-time diarization at 17.5% error across standard benchmarks. Best in class, and one label in six is still wrong.
How does adaptive delay work?
80-millisecond audio chunks, one token each. At every chunk the model chooses to emit a word or ask for more audio, so latency is decided per word. The balance is learned in reinforcement learning with the accuracy and delay rewards multiplied, not added.
Streaming or batch?
Streaming only when words must appear while someone is still talking. On the same datasets the best streaming model is 3.1% and five batch models sit between 1.7% and 2.6%. Picking streaming you did not need costs accuracy and buys nothing.
Sources
- Artificial Analysis — streaming speech-to-text leaderboard, read 1 September 2026. WER, latency and price for every model in the streaming table except GPT Live Transcribe's latency, which is not in the published highlights.
- Artificial Analysis — non-streaming leaderboard, read 1 September 2026. Batch WER figures and dataset weighting.
- The New Stack — Meta just beat OpenAI and Google at real-time transcription, Frederic Lardinois, 1 September 2026. Architecture (80 ms chunks, soft tokens, RL reward), GPT Live Transcribe's 3.9%, the 17.5% diarization figure, the English-only caveat, and the spokesperson's statement on weights.
- 9to5Mac — Meta launches Muse Voice Transcribe, 1 September 2026. Language count, hour-long audio, Mac dictation, pricing.
- 9to5Mac — Meta AI desktop app for Mac, 19 August 2026. Requirements and download.
- OpenAI pricing, read 1 September 2026. GPT Live Transcribe at $0.017 per minute.
- Google Gemini API pricing, read 1 September 2026. Gemini 3.5 Transcribe Live blended rate.
- Deepgram pricing, read 1 September 2026. Flux English streaming at $0.0065 per minute.
- ElevenLabs Speech to Text, read 1 September 2026. Sub-150 ms claim, 90-plus languages, $0.40 per hour.
- Cartesia Ink, read 1 September 2026. Semantic endpointing description and 88 ms turn-detection claim.
- Inworld pricing, read 1 September 2026. STT hourly rates and language count.
Related reading
- Gemini 3.5 Transcribe vs Deepgram vs Whisper — the batch side of the same question
- Streaming transcription on Mac — what streaming changes architecturally, and why MetaWhisp does not use it for dictation
- Meeting transcription without a bot on the call
