Released an open-source EOU detection model covering English + 8 Indic languages voice agents stop cutting callers off mid-sentence

Voice agents serving Indic-language users cut callers off constantly Indic languages put the verb last, so "aapko kal phone…" pauses sound finished to a VAD, and the bot jumps in before "kar dunga" lands. VAD hears silence, not meaning Released an end-of-turn (EOU) detector fine-tuned from Qwen2.5-0.5B scores the live transcript and predicts whether the caller has finished • Languages: English, Hinglish, Hindi, Telugu, Tamil, Marathi, Bengali, Gujarati, Punjabi • CPU-only ONNX Runtime (~2GB, fp32) fully local, no cloud inference • Text-only needs your STT; drop-in for LiveKit Agents, same interface as their text turn detector • Proven live on LiveKit Cloud: held mid-request pauses and committed complete requests 91.3% TPR on our held-out validation set, calibrated to ~10% false cuts on incomplete turns AUC 0.967. On LiveKit's official real-speech harness (Hindi): 85.3% detection at the same budget. Model + plugin: huggingface.co/tarun7r/indic-turn-detector Apache-2.0. Feedback welcome!