Researchers Release TamilEOT Dataset and Models for Voice Agents in Tamil Speech

Researchers have released TamilEOT, an open dataset and model suite designed for semantic end-of-turn detection in Tamil telephone conversations. The dataset contains 18,485 labeled turn boundaries extracted from 116 real telephone calls.

Fine-tuned from Smart Turn v3, the two audio-only models (sized at 8.7 MB and 21 MB) run in under 150 ms single-threaded on a laptop CPU. On a held-out test split of 4,168 clips from 30 unseen calls, detection accuracy improved from a zero-shot baseline of 70.30% to 83.71% and 86.13%, with ROC-AUC reaching 0.921.

The authors noted that labeling using an audio-LLM achieved 97.5% agreement with humans and cost $5.69. The dataset, model weights, code, and training results have been published openly.