BanglaTurn achieves 84.33% accuracy for Bangla end-of-turn detection

BanglaTurn uses a Whisper encoder with task-specific classification heads to detect speech turn states in Bangla. The model reaches 84.33% accuracy on a class-balanced test set, significantly reducing false negative rates from 51.57% to 7.55% with end-to-end CPU latency under 191 ms.