Muse Voice Transcribe: Meta's $0.18/hour real-time diarization for 20+ speakers

Meta's Muse Voice Transcribe is a real-time streaming speech-to-text model that combines transcription, endpoint detection, and speaker diarization for more than 20 speakers in a single model pass, priced at $0.18 per hour of audio processed via the Meta Model API. For enterprise teams building meeting automation, call analytics, or live voice agents, this price-performance combination could shift expectations for what a real-time transcription service should cost and deliver.

What makes Muse different from other STT models

Traditional speech-to-text answers "what was said." Diarization adds "who said it." Muse embeds speaker attribution directly into its autoregressive multimodal architecture rather than running it as a separate post-processing step. Audio arrives in 80-millisecond chunks, and the model uses an adaptive delay mechanism: it can wait longer when speech is ambiguous and commit earlier when context is clear. Meta says reinforcement learning trains this behavior by balancing word-error-rate and delay rewards.

The result is a single model that outputs transcribed text with speaker labels (e.g., <|speaker_A|>) and turn boundaries, all without a separate diarization pipeline. The API provides turn-level timestamps, though not word-level timestamps or confidence scores.

Pricing and competitive positioning

At $0.18 per hour ($3 per 1,000 minutes), Muse is not the absolute cheapest streaming transcription service. Soniox's stt-rt-v5 is $0.12 per hour, but it supports only up to 15 speakers. Muse's pricing includes diarization, while many competitors charge extra for speaker attribution. For example, Deepgram Nova-3 Multilingual costs about $0.35 per hour base plus $0.12 per hour for diarization, totaling roughly $0.47 per hour. AssemblyAI similarly adds $0.12 per hour for streaming diarization on top of its $0.45 base rate.

The table below compares public pricing for one hour of streaming audio with diarization included where applicable:

Muse's $0.18 per hour with diarization included puts it near the low end of the market, especially against providers that charge separately for speaker attribution. At 1,000 hours of audio, Meta's rate implies roughly $180 in transcription costs.

Benchmark claims and caveats

Meta reports that Muse achieved a 3.1% word error rate on the Artificial Analysis AA-WER Streaming Index, ahead of Cartesia Ink-2 (3.4%), ElevenLabs Scribe v2 (3.6%), and GPT Live Transcribe (3.9%). Meta also reports an average diarization error rate of 17.5% across AMI-IHM, AMI-SDM, and VoxConverse, lower than competing systems shown in its chart. These are vendor-reported benchmarks and may not reflect performance in all real-world scenarios.

The 20+ speaker capability is a stated model capacity, not a demonstrated limit in public demos. Meta's principal live demonstration used eight speakers, and its long-form recording had 11 labeled participants. Speechmatics supports up to 100 speakers in real-time, and AWS Transcribe supports up to 30. So Muse's 20+ figure is competitive but not a record.

Deployment considerations

Muse is available via the Meta Model API and is compatible with the OpenAI SDK. It supports long audio exceeding one hour and multilingual code-switching across more than 70 languages, with 25 languages extensively validated initially. The API allows eight concurrent streams per tenant by default and real-time sessions of up to 60 minutes before an application must reconnect. Zero-data-retention processing is priced at parity with standard processing.

Notable limitations: the API provides turn-level timestamps only, not word-level. It does not expose word-level confidence scores, sound-event detection, or emotion detection. Speaker labels are scoped to a session rather than verified identities.

What this means for builders

For teams building meeting intelligence, call analytics, or live voice agents, Muse offers a compelling price-performance point. The integration of diarization into the same model that handles transcription and endpointing reduces pipeline complexity and latency. The main trade-offs are the 20+ speaker ceiling (adequate for most meetings but not large conferences), the lack of word-level timestamps, and the reliance on vendor-reported benchmarks. If your