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

TamilEOT addresses a difficult choice for telephone voice assistants: whether a pause means the caller has finished speaking. A fixed waiting period can interrupt a slow speaker or delay a reply. The project adapts an audio detector to Tamil conversations, reporting accuracy of 86.13% against a 70.30% zero-shot baseline on clips from 30 previously unseen calls. Research abstract.

The surrounding audio pipeline also matters. In a streaming replay, 7.8% of labelled boundaries never reached the detector. Better classification therefore addresses only part of conversational timing: the system must first deliver the relevant audio to the model.

“Without a model of the language that decision falls back to a fixed silence timeout: set it short and the agent interrupts, set it long and every turn pays the full wait.”

“Data, weights, code and every negative result are public.”