TamilEOT: A Dataset and Model for Semantic End-of-Turn Detection in Tamil Telephone Speech

Özetle: A voice agent must decide, at each pause, whether the speaker has finished or is still thinking.

Özet

A voice agent must decide, at each pause, whether the speaker has finished or is still thinking. Without a model of the language, systems fall back to a fixed silence timeout: too short, and the agent interrupts; too long, and every turn pays the full wait. We release TamilEOT, to our knowledge the first open dataset and detector for semantic end-of-turn detection in a South Indian language: 18,485 labelled turn boundaries from 116 real Tamil telephone conversations, two audio-only detectors built on Smart Turn v3's architecture, and a turn-level companion set of 4,774 whole turns with their pause structure. On a held-out split of 4,168 clips from 30 unseen calls, accuracy rises from 70.30% zero-shot to 83.71% (8.7 MB) and 86.13% (21 MB); ROC-AUC rises from 0.751 to 0.921. Both run in under 150 ms single-threaded on a laptop CPU. We also report what building it cost. Our rule-derived negative-class labels scored 44.4% against a blind human pass, below chance, because the rule answered a different question than the model is asked; an audio-LLM labeller at 97.5% agreement replaced them for US$5.69. Of six training levers, only encoder capacity moved the result: three identical runs spanned 0.87 accuracy points, the floor below which nothing else here is a finding. Replaying labelled boundaries through the production VAD and streaming adapter costs a further 2.60 points, and 7.8% of boundaries never reach the model at all. Data, weights, code and every negative result are public.

Orijinal Özet (İngilizce)

A voice agent must decide, at each pause, whether the speaker has finished or is still thinking. Without a model of the language, systems fall back to a fixed silence timeout: too short, and the agent interrupts; too long, and every turn pays the full wait. We release TamilEOT, to our knowledge the first open dataset and detector for semantic end-of-turn detection in a South Indian language: 18,485 labelled turn boundaries from 116 real Tamil telephone conversations, two audio-only detectors built on Smart Turn v3's architecture, and a turn-level companion set of 4,774 whole turns with their pause structure. On a held-out split of 4,168 clips from 30 unseen calls, accuracy rises from 70.30% zero-shot to 83.71% (8.7 MB) and 86.13% (21 MB); ROC-AUC rises from 0.751 to 0.921. Both run in under 150 ms single-threaded on a laptop CPU. We also report what building it cost. Our rule-derived negative-class labels scored 44.4% against a blind human pass, below chance, because the rule answered a different question than the model is asked; an audio-LLM labeller at 97.5% agreement replaced them for US$5.69. Of six training levers, only encoder capacity moved the result: three identical runs spanned 0.87 accuracy points, the floor below which nothing else here is a finding. Replaying labelled boundaries through the production VAD and streaming adapter costs a further 2.60 points, and 7.8% of boundaries never reach the model at all. Data, weights, code and every negative result are public.

Kaynak: arXiv:2609.05631 · PDF