Meta's Superintelligence Labs have released their first real-time audio perception model. It transcribes speech, tells speakers apart, and detects sentence boundaries during live conversation.
The Spark-family model breaks incoming audio into 80-millisecond chunks. After each one, it decides whether to keep listening or output the next word as text, controlling how much context it gathers before committing to a transcription.
More waiting means better accuracy but longer delays. Muse Voice Transcribe adjusts the wait time dynamically for each word based on difficulty, according to Meta. Easy words come out faster, harder ones get more listening time. Meta trained this behavior with reinforcement learning, rewarding the model for low error rates and short delays at the same time.
Adaptive delay shifts the speed-accuracy trade-off because the model decides per word how long to keep listening. | Image: Meta
Speaker separation and sentence detection without extra systems
