Detect Turn Speech (Smart Turn v3)

How it fits

Typical backends

Caveats

• FallbackAn empty input AudioFrame returns probability 1.0 (turn complete) without invoking the model, so silent stretches keep flagging completion — pair with a voice-activity detector upstream if the input may legitimately go quiet.
• State lifetimeThe model only sees the most recent 8 seconds of audio, so conversations with long pauses lose earlier context once the rolling buffer cycles.
• I/O contractMulti-channel input is averaged to mono and anything not already at 16 kHz is resampled every tick; the emitted probability has no internal cutoff, so a downstream gate must apply its own threshold (typically 0.5–0.7).
• Parameter interactionBoth keys are hot-swappable, but changing model, device, or cpu_count mid-stream resets the rolling history and reloads the model on the next tick — do not tune live during a conversation. device `auto` picks CUDA when visible and CPU otherwise, whereas an explicit `cuda`/`cpu` aborts on first inference if that backend is missing.
• Costcpu_count sets the intra-op thread count; values above the host's physical CPU count add scheduling overhead without throughput gain.