Meta has released Muse Voice Transcribe, a real-time speech model that combines transcription, speaker separation and sentence-boundary detection. The system supports more than 70 languages and is available in Meta AI and through the Meta Model API at a listed price of $0.18 per hour.
Incoming audio is divided into 80-millisecond chunks. After each chunk, the model decides whether it has enough context to emit the next word or should wait for more sound. Meta trained that variable delay with reinforcement learning, balancing transcription errors against latency instead of applying one fixed wait to every word.
The same model marks changes among as many as 20 speakers and labels their passages, avoiding a separate diarization pipeline. It also identifies where utterances begin and end. Combining those stages could simplify meeting transcription, call analysis and continuously available voice assistants.
Meta’s accuracy and cost comparisons are vendor claims, and performance will depend on language, accents, background noise and overlapping speech. Continuous-listening applications also create privacy obligations beyond transcription quality. Developers need clear recording indicators, consent, retention limits and controls for sensitive conversations before using a low-latency model as an always-on assistant.