BanglaTurn improves detection of speech turns in Bangla podcasts
BanglaTurn: A Benchmark and Whisper-Based Model for End-of-Turn Detection in Bangla Speech
Computation and Language
Summary
Knowing when one person stops talking and another starts is important for conversations. The authors created BanglaTurn, a collection of short Bangla podcast speech clips labeled to show these turn points. They trained a new model using parts of the Whisper speech recognition system and special classifiers to find turns better than a previous system. Their model is much better at spotting turns that were missed before, and it can work quickly on simpler computers.
What this means in practice
- •For voice assistant developers: Improve conversational flow in Bangla by accurately detecting when a speaker finishes their turn.$Commercial implications: Enables advanced Bangla voice assistants to better manage turn-taking, improving user experience and responsiveness.
- •For call center technology teams: Enhance turn detection in Bangla customer service calls to automate interaction analysis and improve agent training.
Tested on one dataset.
Authors
Mizbaul Haque Maruf
Abstract
This paper presents BanglaTurn, a corpus for end-of-turn detection in Bangla conversational speech, and a model trained on it. The corpus holds 35,374 samples of 3 to 15 s of podcast speech, labelled for turn state by combining speaker diarization with an LLM pass, with every label then checked by a human annotator. The model pairs a Whisper encoder with task-specific classification heads. On a class-balanced test set drawn from a held-out podcast, it reaches 84.33% accuracy (95% CI 80.3 to 88.1) against 69.28% for the Smart-Turn v3 baseline, and lowers the false negative rate from 51.57% to 7.55% at the cost of a higher false positive rate. We report what encoder layer fine-tuning, multi-scale pooling and INT8 quantization each contribute, and latency stays within 165 to 191 ms end to end on CPU.