Meta's first real-time audio perception model. Trained on 70+ languages with diarization for 20+ speakers.