模型與語音 AI
Meta Muse Voice Transcribe Combines ASR, Speaker Diarization, and Endpoint Detection in a Single Streaming Model
Meta’s new speech model outputs transcripts, speaker labels, and utterance boundaries within the same decoding loop, using an adaptive delay mechanism to balance accuracy and latency. It leads English streaming benchmarks and offers low API pricing, but multilingual quality, session limits, and closed weights still complicate deployment decisions.
Meta Superintelligence Labs has released Muse Voice Transcribe, which integrates real-time automatic speech recognition (ASR), speaker diarization, and endpoint detection into a single autoregressive model. Audio is encoded into soft tokens in chunks of approximately 80 milliseconds. The decoder can output text, speaker, and speech-event tokens—or request the next audio segment. Compared with conventional architectures that chain together VAD, ASR, and diarization models, this design can reduce timeline drift between components and lets voice agents decide when to respond directly from endpoint events.
Another core feature is adaptive delay. Rather than applying a fixed wait time to every word, the model determines whether the available acoustic context is sufficient. Easily recognized words can be committed earlier, while ambiguous segments wait for more audio. Meta says the strategy was tuned through reinforcement learning to produce a better Pareto frontier between speed and accuracy. The model was trained on more than 70 languages, with 25 validated at launch. It supports in-sentence code-switching, keyword and contextual biasing, and model-level handling of more than 20 speakers.
On Artificial Analysis’s English-weighted AA-WER Streaming benchmark, the model achieved a final-transcript WER of approximately 3.1% and produced a final result around 0.16 seconds after detecting the end of speech. However, this does not imply equal quality across languages. In independent testing, it outperformed a local Whisper large-v3-turbo deployment on English recordings but performed substantially worse on a single mixed Hindi–English sample. Its speaker and endpoint results were also insufficient to support unsupervised use in high-risk settings.
The service is currently available only through the Meta Model API, Meta AI for Mac, and Muse Code; its weights are not open. Pricing is $3 per 1,000 minutes. Real-time connections through the public API are limited to approximately 60 minutes and cannot be resumed, while the file endpoint imposes even stricter limits. These constraints are not equivalent to the model’s officially demonstrated ability to process more than an hour of audio. Engineering teams should next test Chinese accents, overlapping speech, noise, and disconnect-and-resume behavior, while also confirming transcript revision semantics, data-retention options, and service regions—instead of deciding to migrate solely on the basis of its top ranking in English.