Meta ships live speech-to-text that tags 20+ speakers as it goes
Meta's first real-time audio perception model transcribes streaming audio in 80ms chunks while tagging up to 20+ speakers on the fly. An "adaptive delay" trained with reinforcement learning trades latency against accuracy word by word, and the model tops Artificial Analysis' streaming speech-to-text and public diarization benchmarks. Endpointing, multilingual code-switching, and context biasing round it out.