Build Independently confirmed 2 publishers 3 min read Published Sep 2, 2026, 6:28 PM PDT

Meta folds diarization and endpointing into the transcript stream for $0.18 an hour

Speaker turns and end-of-speech markers now ride in the same token sequence as the words, which buys a cheaper pipeline and a tighter coupling than a three-vendor speech stack. The accuracy claim behind it is English-only.

The Engineer · Build desk

Send a correction

Illustration accompanying Meta folds diarization and endpointing into the transcript stream for $0.18 an hour — Generated illustration

What happened

Meta launched Muse Voice Transcribe on September 1, billing transcription, endpoint detection and speaker diarization from one model at $0.18 per processed audio hour.
The rate is listed as $3 per 1,000 minutes through the Meta Model API, with streaming and non-streaming processing priced the same and diarization included rather than sold as an add-on.
On Artificial Analysis's AA-WER Streaming benchmark, which covers English speech only, Muse reports a 3.1% word error rate against Cartesia Ink-2 at 3.4% and Gemini 3.5 Transcribe Live at 4%.
Muse also leads on speaker recognition at a 17.5% error rate across several standard benchmarks, a band where The New Stack says every comparable model still struggles.
A Meta spokesperson told The New Stack that the weights for this model will not be released, unlike the Muse Glimmer models.

Why it matters

What it costs cost Speech specialists now have to defend their invoices on accuracy in the customer's own languages and on deployment control rather than on rate cards, and the pressure lands hardest on single-purpose diarization and endpointing vendors that have no transcription revenue to cross-subsidise.
What it limits constraint Because speaker turns and end-of-speech markers are emitted inside the same token sequence as the words, there is no seam at which to substitute a better diarizer; you adopt the whole chain or you keep building it yourself.
Who's exposed exposure With no open weights on offer, a team that collapses its speech stack onto this model runs its entire listening path on someone else's endpoint, so a price change or a deprecation notice arrives with no in-house fallback.
What doesn't add up contradiction VentureBeat's read of the launch places the price near the low end while noting specialists beat Meta on individual measures, and Soniox's posted hour is cheaper still, so a migration argued on being the cheapest option is built on the wrong claim.

Next story · Invest · 4d ago

Meta's Muse Wonder, at $20 to $200 a month, is framed as a test of its superintelligence team

Confirmed25 publishers

The unit of purchase here is a token stream. Audio arrives in 80 ms chunks, 12.5 per second, each compressed into one soft token ^[7]. At every chunk the model picks one of two actions: emit a text token, or emit a "next audio" placeholder that the runtime swaps for the following chunk ^[8]. When the audio stops, an "empty audio" token tells the model nothing more is coming, and it flushes whatever text it is still holding ^[9]. Speaker turns ride that same sequence. A `<|start_of_turn|>` token marks a possible change, a token such as `<|speaker_A|>` assigns the turn, and separate tokens mark start and end of speech, all trained alongside transcription rather than handed to a downstream speaker classifier ^[12].

That gives you something to budget with. 12.5 chunks a second across an hour is 45,000 soft tokens ^[22], and $0.18 for that hour works out at $4 per million audio tokens ^[23]. Streaming and batch carry the same rate, and diarization is not a separate line ^[4].

The reward design is the part worth copying. Meta multiplies the word-error-rate reward by the delay reward instead of adding them ^[11]. Addition lets a model buy accuracy with unlimited stalling, because a strong WER term covers a weak delay term. Multiplication makes a near-zero score on either factor collapse the whole product. That is what produces adaptive delay, where the model spends more audio on an ambiguous word and commits early on an easy one ^[10]. What the launch material does not give you is a millisecond figure; the only published timing value is the 80 ms chunk period ^[26]. Barge-in budgets are a distribution here, not a constant, and you will have to measure it yourself.

For the accuracy number to describe your workload, your audio has to be English, streamed, and close enough to whatever mix Artificial Analysis assembled, because that benchmark covers English only ^[13]. Meta says 25 of more than 70 trained languages were extensively validated for this release ^[1], so anything outside those 25 is your eval problem. The margin over the next model on the list is 0.3 points ^[25], in a field where OpenAI, Google, xAI and Alibaba all shipped streaming models within weeks of each other, according to The New Stack ^[21].

The bundled feature is also the weak one. Speaker error at 17.5% is roughly 5.6 times the word error rate ^[24]. If your interface prints a name next to a quote, 17.5% is what the user meets.

On price, the bundle undercuts the $0.24 Speechmatics lists for Standard real-time transcription by 25% ^[19]^[27], while Soniox posts about $0.12 an hour with diarization, language identification and formatting included, a third under Meta ^[18]^[28]. The posted floor belongs to Soniox. What moved is the default assembly: one vendor billing one rate for work that used to need three integrations. Speechmatics will label 50 speakers by default and 100 on request, five times Meta's stated 20-plus ^[19]^[29], and if a hundred people are talking at once, speaker labels are not your first problem. Amazon Transcribe documents 30 ^[20]. Meta is not releasing these weights, a spokesperson told The New Stack ^[15], so a consolidated pipeline is a hosted dependency with no local build behind it.

What to watch

Whether Artificial Analysis publishes a non-English AA-WER Streaming run, since the current 3.1% figure covers English only.
Whether Soniox and Speechmatics reprice or instead defend on speaker counts, language coverage and deployment control.
Whether the $0.18 rate survives once Meta's own glasses and Mac app drive volume, given there are no open weights to fall back on.

Clarity's read

What the record supports and how the coverage leans. The claims behind it follow.

Reality

Evidence55
Adoption24
Hype gap+34
Incentives72
Confidence60

Why these scores

Claim ledger

Ranked by verification strength, evidence, and original report placement.

[1]
Meta's technical launch post says Muse was trained across more than 70 languages, with 25 extensively validated for the first release.
Reported SupportedView cited source
[2]
Meta says Muse supports audio longer than an hour, multilingual code-switching, and more than 20 speakers without a separate diarization pipeline.
Reported SupportedView cited source
[3]
Meta launched Muse Voice Transcribe on September 1, pricing a real-time speech model with transcription, endpoint detection and speaker diarization at $0.18 per processed audio hour.
Reported SupportedView cited source
[4]
Meta lists Muse Voice Transcribe at $3 per 1,000 minutes through the Meta Model API; streaming and non-streaming processing carry the same rate, and speaker diarization is included rather than billed as an add-on.
Reported SupportedView cited source
[5]
At the public rate, 1,000 hours of processed audio would cost $180.
Reported SupportedView cited source
[6]
Meta built Muse Voice Transcribe inside Meta Superintelligence Labs, the AI organization formed after Meta invested $14.3 billion in Scale AI and recruited its founder, Alexandr Wang.
Reported SupportedView cited source
[7]
Audio comes in as 80-millisecond chunks, 12.5 per second, each compressed into a single soft token.
Reported SupportedView cited source
[8]
At every chunk the model either emits a text token or a special "next audio" placeholder, which the system then replaces with the next chunk of audio.
Reported SupportedView cited source
[9]
When the audio stops, an "empty audio" token signals to the model that no more audio is coming, and it flushes any text it is still holding.
Reported SupportedView cited source
[10]
Meta calls the mechanism "adaptive delay": because the model controls how much audio it hears before committing to a word, it controls its own latency, with difficult words getting more context and easy words transcribed almost immediately.
Reported SupportedView cited source
[11]
The tradeoff is learned during the model's reinforcement learning phase, where the word error rate reward and the delay reward are multiplied rather than added.
Reported SupportedView cited source
[12]
Diarization and endpointing are part of the same output sequence: a <|start_of_turn|> token marks a possible speaker change, a token such as <|speaker_A|> assigns the turn, and separate tokens identify the start and end of speech, with those tasks trained together with transcription instead of sending completed text into a separate speaker-classification stage.
Reported SupportedView cited source
[13]
On Artificial Analysis's AA-WER Streaming speech-to-text accuracy benchmark, Muse's word error rate is 3.1%, ahead of Cartesia Ink-2 (3.4%), ElevenLabs Scribe v2 Real-time (3.6%), GPT Live Transcribe (3.9%) and Gemini 3.5 Transcribe Live (4%); the benchmark applies only to English speech.
Reported SupportedView cited source
[14]
On recognizing distinct speakers, all models still struggle more than most users would like, and Muse Voice Transcribe leads with a 17.5% error rate across several standard benchmarks.
Reported SupportedView cited source
[15]
Unlike its Muse Glimmer models, Meta will not make the open weights of Muse Voice Transcribe available, a Meta spokesperson told The New Stack.
Reported SupportedView cited source
[16]
Muse Voice Transcribe is available via the Meta Model API, Meta AI for Mac, and Muse Code.
Reported SupportedView cited source
[17]
A VentureBeat review found that the price puts Muse near the low end of the real-time market, although specialized providers beat Meta on individual measures.
Reported SupportedView cited source
[18]
Soniox lists real-time transcription at roughly $0.12 an hour and includes diarization, language identification and formatting, documenting support for as many as 15 speakers in a session.
Reported SupportedView cited source
[19]
Speechmatics says its real-time service identifies as many as 50 speakers by default and can be configured for 100, with Standard real-time transcription listed at $0.24 an hour.
Reported SupportedView cited source
[20]
Amazon Transcribe documents a maximum of 30 unique speakers and supports speaker partitioning in streaming sessions.
Reported SupportedView cited source
[21]
The New Stack reports that OpenAI, Google, xAI and Alibaba all shipped streaming speech models within weeks of each other, and says a 0.3-point benchmark lead will not hold long in that field.
Reported SupportedView cited source
[22]
One hour of audio at 12.5 chunks per second is 45,000 soft tokens.
Derived Supported
[23]
At $0.18 per audio hour, Muse's soft-token input works out to $4.00 per million tokens.
Derived Supported
[24]
Muse's 17.5% speaker error rate is about 5.6 times its 3.1% word error rate.
Derived Supported
[25]
Muse's word error rate is 0.3 points below the next model on the AA-WER Streaming list, Cartesia Ink-2.
Derived Supported
[26]
None of the supplied reports states a measured latency figure for adaptive delay; the only published timing value is the 80 ms chunk period.
Derived Supported
[27]
Meta's $0.18 hour is 25% below the $0.24 Speechmatics lists for Standard real-time transcription.
Derived Supported
[28]
Soniox's roughly $0.12 hour is about 33% below Meta's $0.18.
Derived Supported
[29]
Speechmatics' configurable ceiling of 100 speakers is five times Meta's stated 20-plus.
Derived Supported

Sources

2 independent publishers whose own reporting we read for this story.

runtimewire.com
1 article · September 2, 2026
Meta launches Muse Voice Transcribe with diarization at $0.18 an hour
thenewstack.io
2 articles · September 1, 2026
Meta just beat OpenAI and Google at real-time transcription

Clarity analysis

How the coverage leans and what the record actually supports. Read the methodology

Perspective Coverage

How much of the coverage speaks to builders, operators and investors. It measures emphasis, not whether the story is right.

Builder

For anyone building on this, the interesting object is the output sequence. Words, speaker turns and speech boundaries share one stream, which means the thing you used to assemble from three vendors is now one decode loop — and one failure surface. The New Stack has the mechanics: 80-millisecond chunks compressed to single soft tokens, an emit-or-consume choice at every chunk, an empty-audio token that flushes held text, and a reward function that multiplies word error against delay so the model cannot buy accuracy with unlimited waiting. Runtime Wire supplies the token names you would actually see. Two practical facts sit at the edges of both reports: the weights are closed, so there is no self-hosted escape hatch, and nobody publishes a millisecond figure for adaptive delay.

Operator

Run this in a conference room and the number that will hurt is 17.5%, not 3.1%. Speaker attribution errs at roughly five and a half times the word rate, and Runtime Wire is the only report explaining why streaming makes it worse: there is little future audio to revise an early speaker guess, and endpointing forces tokens to finalize early. The same piece notes Meta's public demos run eight speakers, with 11 labeled in the long-form recording, against a stated ceiling above 20 — and relays Soniox's and AssemblyAI's advice to test your own noisy rooms, short replies and overlapping voices. Against that, folding three services into one API removes two integrations and, at a flat rate for streaming and batch alike, a lot of billing arithmetic.

Investor

Meta has posted $0.18 an hour with diarization bundled in, and bundling is the aggressive part rather than the number: Soniox is cheaper at about $0.12 and Speechmatics dearer at $0.24 while labeling up to 100 speakers. What Meta is compressing is the price of a whole chain — transcription, turn detection, endpointing, code-switching — which is where speech specialists have priced their margin. Runtime Wire connects the release to the $14.3 billion Scale AI investment that produced Meta Superintelligence Labs and to Zuckerberg's stated intent to distribute personal AI as cheaply as possible, and the glasses and Mac app give Meta a reason to keep pushing whatever the API earns. Set against that: OpenAI, Google, xAI and Alibaba shipped streaming models in the same few weeks, and the accuracy lead being celebrated is 0.3 points wide and English-only.