According to Meta's official announcement, Meta Superintelligence Labs has released the Muse Voice Transcribe model. This model integrates real-time speech-to-text transcription, speaker diarization, and voice endpoint detection into a single architecture, capable of processing audio segments exceeding one hour in length while distinguishing over 20 speakers. In tests conducted by Artificial Analysis for real-time English transcription, the model achieved a word error rate (WER) of 3.1%, outperforming GPT Live Transcribe at 3.9% and Gemini 3.5 Transcribe Live at 4.0%.

The model delivers final text approximately 0.16 seconds after a speaker finishes speaking, employing a dynamic waiting strategy: it processes 80ms audio chunks at a time and autonomously decides whether to continue listening or output results—simple words are delivered faster, while more complex words wait for additional context. Meta leverages reinforcement learning to simultaneously optimize accuracy and latency. Trained on over 70 languages—with 25 languages receiving focused validation—the model supports mixed-language speech recognition, including Chinese-English bilingual input. Muse Voice Transcribe is now integrated into Meta AI for Mac and Muse Code, and the Meta Model API is publicly available, priced at $3 per 1,000 minutes ($0.18 per hour).