Select application-specific operating thresholds

Determine decision thresholds on held-out data for specific end-of-turn detection applications so that the trade-off between missed endpoints and false interruptions is appropriate to each application.

Background

The BanglaTurn model produces endpoint probabilities, allowing its operating threshold to be adjusted. At the reported threshold of 0.5, the model achieves a low false-negative rate but produces more false positives than Smart-Turn v3, creating a risk of interrupting speakers. The appropriate balance depends on the application, such as a fast-responding assistant, dictation, or counselling, but the paper evaluates only the 0.5 operating point.