Meta Superintelligence Labs has released Muse Voice Transcribe, a streaming speech model that identifies who is talking inside the same network that writes down what they said, priced at $3 per 1,000 minutes — $0.18 an hour, diarization included, streaming and non-streaming billed identically. Meta claims support for more than 20 speakers in a single session. That is not the highest published limit on the market, and it is not what makes the launch matter. The price is roughly a third of what the nearest comparable vendors charge once their separate diarization fees are added in.
Diarization is the second question a transcript has to answer. Speech recognition handles "what was said"; diarization handles "who said it". The distinction turns into money the moment a transcript is handed to another AI system rather than to a human. A meeting assistant can get every sentence right and still produce a worthless record if it attributes an approval, a commitment or an objection to the wrong participant. The same failure shows up in support analytics, compliance workflows and any agent operating in a room with more than one person in it.
Most vendors solve this with a clustering pass after transcription. Meta built it into the token stream. Audio arrives in 80-millisecond blocks — 12.5 per second — each turned into a soft token, and at every step the model chooses between taking more audio and emitting text. Meta calls this adaptive latency: wait longer when the speech is ambiguous, commit earlier when the context is sufficient. The behaviour was trained with reinforcement learning, combining a reward for word error rate with a reward for latency. Speaker identity and turn boundaries live in the same sequence: one token marks a possible start of a new turn, others indicate the speaker, and separate start and end tokens fix the edges of speech. Recognition, diarization and endpointing are trained together instead of being stitched together afterwards.
The speaker-count claim needs deflating, and Meta's own materials do most of the deflating. Speechmatics supports real-time diarization with 50 speakers by default and up to 100 on a raised limit. Amazon Transcribe distinguishes up to 30 unique speakers, streaming included. Soniox does real-time and batch diarization with a maximum of 15. AssemblyAI lets you set max_speakers between one and 10. The xAI Speech-to-Text API supports streaming diarization without publishing a maximum in its documentation. Comparing these numbers directly is already shaky — vendors implement diarization differently and not all of them publish a ceiling — but the highest explicitly published limit among them belongs to Speechmatics, not Meta. And Meta did not demonstrate 20-plus speakers at launch: the main demo runs with eight, and the long annotated recording has 11. "More than 20" is a stated capability, not something anyone has watched happen.
The pricing is where the announcement has teeth. Billing counts only audio actually processed, rounded down to whole seconds, and Meta says zero-retention processing costs the same as the standard mode — a detail that matters more to regulated buyers than any benchmark on the page. Against published rates: Qwen's international real-time tier runs $0.00009 per second, about $0.324 an hour. AWS varies by region and volume, with streaming in the Northern Virginia example around $0.60 an hour. ElevenLabs lists $0.39 an hour on API pricing, $0.28 or less on annual enterprise plans. Deepgram's Nova-3 Multilingual streaming is roughly $0.35 an hour with diarization adding $0.002 a minute, about $0.47 all in. AssemblyAI is $0.45 an hour for Universal-3.5 Pro Realtime plus another $0.12 for streaming diarization. Google's Gemini figure is a blended token estimate rather than a fixed hourly rate, and Cartesia's Ink-2 is sold as monthly credit packs — the $5 Pro plan covers about nine hours and 16 minutes, roughly $0.54 an hour if every credit goes to transcription, which is a derived number and not an hourly API rate.
Muse is not the cheapest streaming transcription on the market; Soniox currently publishes a lower equivalent price. It is the cheapest one that includes speaker attribution in the base rate. A thousand hours of audio costs about $180 under Meta's public pricing.
That number is the tell. $180 per 1,000 hours is not a revenue line for a company of Meta's size, and the launch reads less like a product with a P&L than like a decision to make speaker-attributed streaming transcription cost nothing in particular. Vendors charging a separate diarization fee now have to explain what that fee buys, and the ones whose entire pitch was competitive per-hour pricing have to compete on accuracy instead. That is an uncomfortable place for most of this field to be pushed.
Which brings up the accuracy Meta chose to lead with, and the accuracy it reported more quietly. On the Artificial Analysis AA-WER streaming index, Muse scores a 3.1% word error rate, ahead of Cartesia Ink-2 at 3.4%, ElevenLabs Scribe v2 Realtime at 3.6%, Qwen3 ASR Flash Realtime at 3.7%, GPT Live Transcribe and Grok Speech to Text Streaming at 3.9%, and Gemini 3.5 Transcribe Live and AssemblyAI U3.5 Realtime Pro at 4.0%. Meta states that as of September 1 this put Muse first on the streaming benchmark run by Artificial Analysis, an independent third-party evaluator. Separately, Meta reports an average diarization error of 17.5% across AMI-IHM, AMI-SDM and VoxConverse, which it says is lower than the competing systems on its own chart.
Word error rate and diarization error are different metrics and cannot be divided into each other, but the gap is worth sitting with: the number describing what was said is best-in-class, and the number describing who said it is more than five times larger, in a product whose distinguishing feature is knowing who said it. Meta's diarization chart also does not test competitors at their claimed maximum speaker counts, and speaker capacity is not a proxy for attribution quality — a system that can label 100 participants is not thereby better at assigning a given sentence than one capped at 20.
The constraints buyers will hit first are the ones sitting quietly in the documentation. Real-time sessions are capped at 60 minutes, after which the application has to reconnect — against a launch claim of support for audio longer than an hour. Eight concurrent streams per customer by default. Timestamps are emitted for turns, not for individual words. There are no per-word confidence scores, no audio event detection and no emotion detection. Speaker labels like A and B are scoped to the session and assert nothing about identity. For a meeting-notes product those are inconveniences. For a pipeline that feeds transcripts to an agent that then acts on them, the absence of per-word confidence is the sharp edge: there is no signal to gate on, no way to flag the span where attribution is shaky before a wrongly assigned commitment travels downstream as fact.
Meta has made speaker-attributed streaming transcription cost 18 cents an hour. The 17.5% figure says the thing being turned into a commodity is not finished yet, and the buyers most attracted by the price are the ones building systems that act on the output without a human reading it first.