Today, Fable 5.1 and World Labs’ Atlas were released, and everyone is busy checking how much smarter the models have become and how far world models have advanced. In a nearly overlooked corner, Meta released something I think is well worth discussing: Muse Voice Transcribe, the first real-time audio perception model from Meta Superintelligence Labs. On the surface, it’s just a speech-to-text model—but it does far more than traditional ASR systems like Whisper. Real-time transcription, speaker diarization (identifying who is speaking at any given moment), and endpointing (determining when a speaker has finished) are all natively performed within a single streaming model. No need to stitch together separate ASR, VAD, and speaker identification components into a pipeline. This is actually significant. A major pain point in voice agents today isn’t just transcription accuracy—it’s knowing when to stop talking and listen, when it’s their turn to speak, and—who among multiple people in a room—is actually addressing them. Muse’s approach is also intriguing. Audio is chunked every 80ms, and the model decides autonomously whether to keep listening or begin outputting text. Meta further trained an adaptive delay using RL: simple words are confirmed quickly; complex ones wait for more context. The results are impressive. On Artificial Analysis’s streaming speech recognition leaderboard, Muse Voice Transcribe achieves a 3.1% WER with an end-to-end latency of approximately 160ms—currently the best. It was trained on over 70 languages, with 25 verified at launch, and can seamlessly handle code-switching between Chinese and English within a single utterance. It processes over an hour of audio with more than 20 speakers without requiring additional post-processing. And it costs just $0.18 per hour. I find this to be a quintessentially Meta release. While everyone else has been racing to boost AI intelligence over the past year, Meta is now focusing on restoring its senses.