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

> Retained reddit record from reddit:LocalLLaMA.

Canonical: https://highsignal.app/data/records/7cedda97808d3e9d

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:LocalLLaMA)  published 2026-11-07 16:08 ingested 2026-11-08 02:39

[original source ↗](https://www.reddit.com/r/LocalLLaMA/comments/1xq9yjf/released_an_opensource_eou_detection_model/)

## retained text



```
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: https://huggingface.co/tarun7r/indic-turn-detector

Apache-2.0. Feedback welcome!
```



document fetched 2026-11-08 00:34 · canonical https://reddit.com/r/LocalLLaMA/comments/1xq9yjf/released_an_opensource_eou_detection_model
