Aayush523478950085/dual-signal-turn-detection
0
Dual-Signal Turn Detection Model
This HuggingFace Space hosts the interactive demo for the Dual-Signal Turn Detection Model designed for low-latency conversational voice AI.
Note on Latency: Live mic latency on HuggingFace Spaces reflects Spaces CPU execution — local benchmark is 8.3ms on CPU.
Architecture
- Fast Path (Stage 1): 19 acoustic features extracted via
librosa.yin(<3ms) + LightGBM classifier. - Slow Path (Stage 2): Frozen
openai/whisper-tinyencoder + trailing pooling + 2-layer MLP head. - Cascaded Gating:
TURN_GATE = 0.82,HOLD_GATE = 0.20, Stage 3 Double-Uncertainty Fast Path rescue.
