TamilEOT releases open models for detecting when Tamil callers finish speaking

Researchers released TamilEOT, an open dataset and two models that detect whether a Tamil telephone caller has finished speaking or is pausing. Their preprint reports accuracy of 83.71% and 86.13%, compared with 70.30% without task-specific training, on 4,168 clips from 30 unseen calls. Both models run in under 150 ms on a single laptop CPU thread. The dataset contains 18,485 labelled turn boundaries from 116 conversations; data, model weights and code are public. Tests through the production voice-activity detector and streaming adapter reduced accuracy by a further 2.60 points, while 7.8% of boundaries never reached the model.