Meta launches Muse Voice Transcribe with diarization at $0.18 an hour

Why it matters

Meta is pricing transcription, endpointing and speaker diarization as one low-cost API, pressuring speech specialists to defend their rates with better accuracy, deployment controls and production reliability.

Mark Zuckerberg's Meta launched Muse Voice Transcribe on September 1st, pricing a real-time speech model with transcription, endpoint detection and speaker diarization at $0.18 per processed audio hour. The release gives developers one model for hearing words, separating speakers and deciding when a person has finished talking, three jobs that voice applications often handle through separate systems.

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. Wang started Scale in 2016 and spent years building the data infrastructure used to train other AI developers' models. At Meta, the assignment is broader: build models that Meta can place across its own products and sell to developers through an API.

Zuckerberg, who founded Facebook in 2004 after studying computer science at Harvard, still sets Meta's product and technology strategy. His August essay, "The Future is for Everyone", committed Meta to distributing personal AI widely and offering it free or as affordably as possible. Muse Voice Transcribe turns that language into a concrete infrastructure bet. Cheap, speaker-aware listening gives Meta a path into meeting tools, call analytics, dictation software and voice agents built by developers outside its social apps.

Meta's technical launch post says Muse was trained across more than 70 languages, with 25 extensively validated for the first release. Meta says it supports audio longer than an hour, multilingual code-switching and more than 20 speakers without a separate diarization pipeline. 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.

The cheap part is the stack

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, according to the developer material reviewed by VentureBeat, while speaker diarization is included rather than billed as an add-on. At the public rate, 1,000 hours of processed audio would cost $180.

Meta does not have the lowest posted price. Soniox lists real-time transcription at roughly $0.12 an hour and includes diarization, language identification and formatting. Soniox documents support for as many as 15 speakers in a session.

Meta also does not have the largest published speaker count. 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. Amazon Transcribe documents a maximum of 30 unique speakers and supports speaker partitioning in streaming sessions.

Muse's commercial case rests on its combination of speaker capacity, integrated voice features and price, rather than on a speaker-count record. Meta combines 20-plus-speaker attribution, live transcription, turn detection, code-switching and context biasing at one rate. A developer building a meeting assistant or an agent that listens in a room may care more about the cost and behavior of that complete chain than the maximum number of speaker labels available in an unusual edge case.

Meta trained the latency decision

Muse receives audio in 80-millisecond chunks, converting each chunk into a soft token. The model then chooses whether to consume another piece of audio or emit text. Meta calls the mechanism "adaptive delay": Muse can wait for context when a word is ambiguous and answer sooner when the available audio is sufficient.

Meta says reinforcement learning balances word-error-rate and delay rewards during training. That matters for voice software because a uniformly fast transcript can commit to the wrong word, while a uniformly cautious one makes an agent feel slow. Muse attempts to vary that trade-off word by word.

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. Meta trains those tasks with transcription instead of sending completed text into a separate speaker-clustering service.

That architecture also fits Meta's own hardware strategy. A pair of AI glasses or an ambient assistant needs to distinguish its wearer from other people, handle interruptions and recognize when a request is complete. Meta says the model is available through Meta AI for Mac and Muse Code, and through its API for other applications.

Claims meet conference rooms

Meta's public launch material describes Muse's capabilities, but the supplied materials do not establish independent word-error-rate, diarization-error or latency results. Production performance will depend on microphones, accents, background noise, short replies and overlapping speech.

The public demonstrations also stop short of the headline capacity. Meta's main live demo uses eight speakers, while its long-form recording contains 11 labeled participants, according to VentureBeat's review. The 20-plus figure is a stated capability rather than a stress test shown in the launch material.

Streaming diarization is especially sensitive to real-world conditions because the system has limited future audio to correct an early speaker assignment. Soniox warns that attribution errors can rise in real time and that endpoint detection can reduce diarization accuracy by forcing tokens to finalize early. AssemblyAI advises customers to test short responses, noisy rooms and overlapping voices against their own recordings.

Those tests matter because a meeting assistant can transcribe every sentence correctly and still create an unreliable record if it assigns an approval, commitment or objection to the wrong participant. The same failure can contaminate call analytics and confuse voice agents operating around several people. Meta has not publicly established enterprise terms covering service guarantees, regional availability or other production controls.

The production test is still ahead

Meta can use Muse inside its apps, AI glasses and developer tools, then expose the same capability as a metered service. The $0.18 rate lowers the cost of testing an integrated transcription stack and gives developers an alternative to assembling endpointing and speaker attribution from separate providers.

Specialist speech providers retain clear advantages. Soniox posts a lower rate, Speechmatics supports more speakers, and Mistral's Voxtral Realtime offers open weights under Apache 2.0 with latency configurable below 200 milliseconds. Enterprise buyers will also test retention controls, regional deployment, concurrency and service guarantees before replacing a production transcription provider.

Meta has set an aggressive reference price for an integrated voice layer. Founders can now test speaker-aware speech software without paying separately for every perception step, but price does not settle the harder questions around attribution, latency and reliability. Muse's practical test begins once 20 people stop reading a script and start interrupting one another.