Most production voice systems are built by stitching together three separate components: one model handles transcription, a second distinguishes speakers, and a third determines when a user has finished speaking. Each transition between these components adds latency and introduces new opportunities for failure.
This week, Meta Superintelligence Labs introduced Muse Voice Transcribe, a single autoregressive model that unifies these three tasks. Described as Meta's first real-time audio perception model, Muse Voice Transcribe performs streaming automatic speech recognition (ASR), speaker diarization for up to 20+ speakers, and endpointing in a single passβwithout requiring any post-processing.
Deployment and Availability
Muse Voice Transcribe is available exclusively as a hosted API, with no open-source weights released. It is currently live on the Meta Model API under the model identifier muse-voice-transcribe-1.0, priced at $3.00 per 1,000 audio minutes (equivalent to $0.18 per hour). The model already powers dictation features in Meta AI for Mac and Muse Code.
Technical Foundation: Streaming ASR
Muse Voice Transcribe belongs to the Muse Spark family of autoregressive multimodal models. Audio is processed in 80-millisecond chunks at a rate of 12.5 Hz, with each chunk transformed into a single soft token. After processing each chunk, the model makes a binary decision: it either generates a <|next_audio|> token to continue listening or emits a transcript token to produce text output.
Looking ahead to 2026, this architecture positions Muse Voice Transcribe as a significant step toward more cohesive, low-latency voice interfaces. By eliminating the traditional pipeline delays, it offers a glimpse into how future AI systems might handle real-time conversational interactions with greater fluidity and accuracy.
via MarkTechPost