Shiprocket Turn Detection
Audio-native complete/incomplete classifier built for the Shiprocket assignment.
Intended use
Run after VAD detects a pause. Input: up to 8 seconds of 16 kHz mono audio. Output: P(complete).
Training data
Only pipecat-ai/smart-turn-data-v3.2-train. Indic results use official Hindi/Marathi slices.
Limitations
- Not a full-duplex conversational model.
- Clip metrics do not prove live interruption quality.
- Browser preprocessing is approximate; use the Python path as reference.
- Do not use for safety-critical barge-in.
Metrics
Fill after the first trained export. Until then the demo may fall back to Pipecat Smart Turn v3.2.