Meta launched Muse Voice Transcribe at $0.18 per hour of processed audio, making it one of the most aggressively priced real-time speech-to-text APIs on the market. Built by Meta Superintelligence Labs, the model handles streaming transcription, endpoint detection, and speaker diarization for more than 20 speakers in a single unified architecture, trained across 70 languages with 25 extensively validated at launch. Streaming and non-streaming transcription are priced identically, and zero-data-retention processing costs the same as standard processing.
The 20-speaker figure is not a record. Speechmatics documents a default ceiling of 50 speakers in real-time, configurable up to 100. Amazon Transcribe supports 30 unique speakers in streaming mode. Meta's own launch demos show 8 speakers live and 11 in a long-form recording sample. What Muse does differently is architectural: audio arrives in 80-millisecond chunks, each converted to a soft token, and the model uses reinforcement learning to balance word-error-rate against latency, committing output early when context is sufficient and waiting longer when speech is ambiguous. Speaker labels, turn boundaries, and endpoint tokens are all emitted in the same autoregressive token sequence, not handed off to a separate post-processing pipeline.
The full article is worth reading for the competitive pricing table across vendors and the technical breakdown of how Muse's adaptive delay mechanism works at the token level. The core question it raises is whether unified architecture plus sub-$0.20 per hour pricing changes the build-vs-buy calculation for enterprise developers in meeting analytics, compliance, and ambient AI, even if the speaker-count ceiling lags Speechmatics.
[READ ORIGINAL →]