Meta Prices Muse Voice Transcribe at $0.18 an Hour, with Real-Time Diarization for 20+ Speakers: A Steal for Enterprises? | VentureBeat
The Independent Voice on AI Orchestration Infrastructure Data Security Technology More Newsletters
Meta is entering the increasingly competitive real-time speech-to-text market with Muse Voice Transcribe, a new audio perception model that combines streaming transcription, endpoint detection, and speaker diarization for more than 20 speakers—at a public API price of just $0.18 per hour of processed audio.
Developed by Meta Superintelligence Labs, Muse is designed to process speech while it happens rather than waiting for a recording to finish. Meta’s launch post for Muse Voice Transcribe states the model supports long audio exceeding an hour, seamless multilingual code-switching, language and keyword biasing, and diarization without a separate post-processing pipeline. The model was trained across more than 70 languages, with 25 extensively validated for the initial release.
The 20-plus-speaker figure is substantial, but it is not a world record. A review of current vendor documentation turns up systems with higher published ceilings:
- Speechmatics' real-time transcription service claims it can identify 50 speakers by default and up to 100 when the limit is increased.
- Amazon Transcribe's diarization documentation specifies a maximum of 30 unique speakers, including for streaming transcription.
Muse nevertheless lands toward the high end of the market, and Meta's broader proposition is arguably more important than the raw maximum: high-capacity real-time diarization combined with low-latency transcription, endpointing, multilingual code-switching, and aggressive API pricing in the same model.
For enterprise developers building meeting systems, call analytics, live assistants, or ambient AI, that combination could matter more than who holds the speaker-count record.
Diarization is becoming part of the core voice stack
Traditional speech recognition answers the relatively simple question, "What was said?" Diarization adds another layer: "Who said it?" This distinction becomes critical as transcripts feed downstream AI systems. A meeting assistant can correctly transcribe every sentence and still create an unreliable corporate record if it attributes an approval, commitment, or objection to the wrong participant. The same issue affects customer service analytics, compliance workflows, and AI agents operating in rooms where several people can speak.
Muse incorporates speaker attribution directly into its autoregressive multimodal architecture. Meta explains that audio arrives in 80-millisecond chunks (12.5 chunks per second), each transformed into a soft token. At each step, the model decides whether to consume more audio or emit text. Meta calls this mechanism adaptive delay: rather than applying one latency budget to every word, Muse can wait longer when speech is ambiguous and commit earlier when it has enough context. Meta states that reinforcement learning combines word-error-rate and delay rewards to train this behavior.
Meta's technical explanation of Muse details the architecture.
Speaker attribution and endpointing then become part of the same token sequence. A <|start_of_turn|> token marks a potential new speaker turn, tokens such as <|speaker_A|> identify the speaker, and separate onset and endpoint tokens identify speech boundaries. Meta says it trains ASR, diarization, and endpointing together rather than running speaker clustering as an unrelated downstream process.
Meta's Model API speech-to-text documentation also exposes diarization as a first-class operating mode alongside push-to-talk and endpointing. Speaker labels such as