Bangla speech model improves detection of when speakers finish

Researchers report that a model trained to detect when a Bangla speaker has finished talking reached 84.33% accuracy on a held-out podcast test set, compared with 69.28% for the Smart-Turn v3 baseline. Its false negative rate fell from 51.57% to 7.55%, which could help a voice system avoid missing the moment to respond. The model uses a Whisper speech encoder with classification components trained on 35,374 labelled podcast samples. Its measured processing time was 165 to 191 milliseconds on a CPU. The lower false negative rate came with a higher false positive rate, meaning the system more often risks treating an unfinished turn as complete.