Blame for README.md

---
license: cc-by-4.0
language: [hi, mr, ta, kn, ml, gu, pa, te, as, or, bn]
task_categories: [audio-classification]
pretty_name: Indic Smart Turn data
---

# Indic Smart Turn data

Turn-boundary labels for training an Indic end-of-turn detector with the [Pipecat Smart Turn v3](https://github.com/pipecat-ai/smart-turn).

Source audio is [ai4bharat/IndicVoices](https://huggingface.co/datasets/ai4bharat/IndicVoices) (CC BY 4.0), conversational rows only: one speaker's side of real two-party phone calls, split into transcript segments.

## Files

- `index/<lang>.parquet`: conversational rows selected per language (shard, row, session, chunk, speaker, duration, text). No audio.
- `labels/<lang>.jsonl`: text label per segment end, from an LLM reading the ordered transcript of a session (`endpoint_bool`, `llm_conf`, `label_model`).
- `labels/<lang>.audio.jsonl`: Gemini audio verdict per built clip (`verdict`, `confidence`, `reason`), keyed by `session|chunk|dataset|md5(audio)[:10]`. This is the primary label. Includes verdicts for pause-cut clips (`dataset` ends in `_pausecut`).
- `built/<lang>.parquet`: 8 s 16 kHz WAV clips (`audio` struct of bytes) with `endpoint_bool` (Gemini audio verdict, primary), `text_label`, `audio_conf`, `audio_reason`, `ambiguous`, `spoken_text`, `speaker_id`, `session`, `chunk`, and a speaker-disjoint `split` (`train`, `dev`, `test`, `test_ambiguous`). Pause-cut rows (`dataset` ends in `_pausecut`) are kept only when the audio verdict is incomplete. 50,421 samples in total.

Languages: Hindi, Marathi, Tamil, Kannada, Malayalam, Gujarati, Punjabi, Telugu, Assamese, Odia, Bengali (ISO-639-3 codes in file names).

## How labels were made

See the [project repository](https://github.com/adimyth/indic-smart-turn) for the full method and code. Short version: a text LLM reads each session's transcript and labels every segment end complete/incomplete; `gemini-3.7-flash` then hears every built clip and gives the deciding verdict. Text and audio agree on 89.5–93.3% of segments per language; only 39–50% of pause-cut clips are truly incomplete, so pause-cuts are kept only when the audio verdict says incomplete.

Attribution: AI4Bharat (IndicVoices), SPRING Lab IIT Madras / TamilEOT, Pipecat.