Meta 推出 Muse Voice Transcribe，为 MSL 首个实时音频感知模型，今日开始推送，自适应延迟下最终转写词错率低至 3.1%。该模型以 80ms 分块处理音频，在单一流式模型内原生处理说话人分离和端点检测，可自行判断何时输出文字、说话人何时切换；通过 Meta Model API、Meta AI for Mac 和 Muse Code 提供。
Love this, another huge release from Meta.
Lunched Muse Voice Transcribe for real-time voice dictation, with the lowest 3.1% final-transcription word error rate with adaptive delay.
which is significantly ahead of competing models.
- The big deal is that Muse does something beyond just transcribing speech; it learns when to wait, when to commit a word, when a speaker changes, and when a turn is actually over, all inside the same streaming model. That makes it much closer to a real-time perception layer for voice agents than a conventional speech-to-text API.
- Muse processes audio in 80ms chunks and chooses after each chunk whether to emit text or keep listening.
Meta made it available through Meta Model API, Meta AI for Mac, and Muse Code
Muse Voice Transcribe is MSL's first real-time audio perception model -- rolling out today. SOTA in streaming speech-to-text, it handles speaker diarization, and endpointing natively in a single model.在 X 查看被引用的帖子
来源：Rohan Paul · x.com