README.md

---
license: cc-by-4.0
language: [ta]
task_categories: [audio-classification]
tags: [end-of-turn, turn-detection, turn-taking, endpointing, voice-agent, tamil, south-indian, smart-turn]
size_categories: [10K<n<100K]
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: dev
    path: data/dev-*
  - split: test
    path: data/test-*
dataset_info:
  features:
  - name: audio
    dtype: audio
  - name: endpoint_bool
    dtype: bool
  - name: language
    dtype: string
  - name: dataset
    dtype: string
  - name: sid
    dtype: string
  - name: call
    dtype: string
  - name: split
    dtype: string
  - name: source
    dtype: string
  - name: harvest
    dtype: string
  - name: gap
    dtype: float32
  - name: prev_dur
    dtype: float32
  - name: n_samples
    dtype: int32
  - name: label_pipeline
    dtype: bool
  - name: llm_verdict
    dtype: bool
  - name: dispute
    dtype: bool
  - name: llm_conf
    dtype: float32
  splits:
  - name: train
    num_examples: 11992
  - name: dev
    num_examples: 2325
  - name: test
    num_examples: 4168
---

# tamil-eot

**18,485 labelled end-of-turn boundaries from 116 Tamil telephone conversations.**

Each row is up to 8 seconds of one speaker's audio ending at a turn boundary,
labelled `complete` (the speaker finished) or `incomplete` (they paused
mid-thought). Built for training semantic turn detectors for voice agents;
Tamil is not covered by Smart Turn v3 or the LiveKit turn detector.

Schema matches `pipecat-ai`'s Smart Turn datasets, so it trains with their loop
unchanged.

| | |
|---|---|
| clips | 18,485 — 16 kHz mono FLAC, ≤ 8 s |
| complete / incomplete | 12,069 / 6,416 |
| source calls | 116 |
| train / dev / test | 11,992 / 2,325 / 4,168 |
| source audio | ~36 h |

## Structure

**Split by call, not by clip** — no call appears in two splits. The test split
is fixed and has not changed across releases.

| field | meaning |
|---|---|
| `audio` | 16 kHz mono FLAC |
| `endpoint_bool` | **the label** — true = speaker finished |
| `language` | `tam` |
| `dataset` | provenance tag |
| `sid` | stable clip id, `<call>_<L\|R>_<offset_ms>` |
| `call` | source call — **group by this** |
| `source` | boundary type: `change`, `change_midseg`, `hold_intra`, `hold_inter` |
| `harvest` | `core`, or `relaxed` for the widened-gate second pass |
| `gap` | silence at the boundary, seconds |
| `prev_dur` | preceding talk-spurt duration, seconds |
| `n_samples` | true length before padding |
| `label_pipeline` | first-pass acoustic label |
| `llm_verdict` | audio-LLM label |
| `dispute` | the two disagree (7,471 rows) |
| `llm_conf` | labeller confidence |

`source` distinguishes speaker handovers (`change`, `change_midseg`) from
same-speaker pauses (`hold_intra`, `hold_inter`). Handovers are ~93%
single-class; `hold_intra` is near-balanced and is 58% of the test split.

```python
from datasets import load_dataset
ds = load_dataset("santhosh-005/tamil-eot")
ds["train"][0]["audio"], ds["train"][0]["endpoint_bool"]
```

`datasets` ≥ 4 needs `torchcodec` to decode the `audio` column. To avoid that
dependency, read the FLAC bytes directly:

```python
import io, soundfile as sf
from datasets import load_dataset, Audio

ds = load_dataset("santhosh-005/tamil-eot").cast_column("audio", Audio(decode=False))
x, sr = sf.read(io.BytesIO(ds["train"][0]["audio"]["bytes"]))
```

Trained on this: [`santhosh-005/smart-turn-tamil`](https://huggingface.co/santhosh-005/smart-turn-tamil)
— 83.71% (tiny) and 86.13% (base) on the held-out test split, against 70.30%
zero-shot. Usable from LiveKit Agents via
[`smart-turn-livekit`](https://pypi.org/project/smart-turn-livekit/).

## Dataset creation

### Source

Derived from **SPRING_INX Tamil R1** (SPRING Lab, IIT Madras, CC BY 4.0), which
ships each call as **one audio file per speaker**, separately transcribed. That
makes the speaker label structural rather than inferred — no diarization, and
no speaker-error rate to propagate.

### Boundaries and cutting

Turn boundaries come from Silero VAD run per leg, not from transcript
timestamps: the shipped segments are padded and sum to ~107% of call
wall-clock, so their edges are not turn boundaries. Transcripts are used only
at build time — filtering crosstalk bleed, counting words for the backchannel
gate, and separating mid-segment boundaries — and are never part of a clip.

Of 50,532 candidate boundaries, **18,485 were kept**. Gates require ≥1.0 s of
speech ending at the cut (enough prosody to read), ≤0.05 s of crosstalk in the
gap, handovers within 1.5 s followed by ≥3 words (rejecting backchannels such
as *um*, *okay*), and pauses between 0.2 s and 2.0 s. A second pass widened the
speech-duration and pause limits to 0.5 s and 5.0 s, adding 2,272 clips marked
`harvest = "relaxed"`.

Each clip is cut from **one leg only** — what a deployed detector receives is
the inbound user stream, not a mixdown — and ends with **200 ms of real audio
after the boundary, identical for both classes**. The 0.2 s pause floor exists
because the speaker who paused is the one who resumes: a shorter pause would
place their resumed speech inside the clip. Measured trailing-window peak
amplitude is 0.0021 median, so trailing-silence length carries no label
information.

### Labels

The first labelling pass derived labels from VAD boundaries plus transcript
segment structure — a pause falling inside a transcript segment was treated as
`incomplete`. Audited against a human listening to 197 clips blind, it scored
**70.1% overall and 44.4% on same-speaker pauses**, below chance, because
"a pause fell inside a transcript segment" is not a judgement about whether a
thought was finished.

Labels were rebuilt with an **audio LLM listening to the clip only** — no
transcript, no sight of the first-pass label. Seven candidates were scored
against the same 197 human-labelled clips before any were used:

| labeller | agreement | `hold_intra` | precision on `incomplete` |
|---|---|---|---|
| **`gemini-3.7-flash`** *(used)* | **97.5%** | 97.0% | 93.9% |
| `gemini-3.6-flash` | 96.4% | 97.0% | 88.7% |
| `gemini-3-flash-preview` | 93.4% | 90.9% | 85.7% |
| first-pass acoustic | 70.1% | 44.4% | 44.4% |

`gemini-3.7-flash` tied statistically with `gemini-3.1-pro`; price broke the
tie. Labelling the full set cost $5.69. Relabelling flipped 53% of the original
negatives and removed a confound where the gates had drawn negatives from
longer utterances (Cohen's *d* on `prev_dur`: −0.33 → −0.09).

Both labels ship on every row. `endpoint_bool` follows `llm_verdict`.

## Considerations

- **One domain** — narrowband two-party telephone conversations, task-oriented.
- **Positive-heavy** at 65:35, not 50:50.
- **Label-noise ceiling ~97.1%.** Above that you are fitting labeller error.
- **`source` is unbalanced.** Report per-`source` accuracy against each
  bucket's own base rate; overall accuracy is dominated by the near
  single-class handover buckets.
- **Human reference labels are not gold** — one pass by ear on isolated clips,
  enough to rank labellers, not authoritative on any single clip.
- Audio is conversational speech from a public corpus; no personally
  identifying metadata is included beyond what SPRING_INX ships.

## Licensing and citation

The **dataset** — boundary extraction, cut geometry, gates, labels, and splits
— is by [santhosh-005](https://github.com/santhosh-005), released under
**CC BY 4.0**.

```bibtex
@misc{tamil-eot,
  author = {santhosh-005},
  title  = {tamil-eot: labelled end-of-turn boundaries for Tamil},
  year   = {2026},
  url    = {https://github.com/santhosh-005/tamil-eot}
}
```

The **source audio** is from SPRING_INX Tamil R1 and remains the work of SPRING
Lab, IIT Madras. It is redistributable here under CC BY 4.0, with attribution:

> SPRING Lab, Indian Institute of Technology Madras. *SPRING_INX Tamil R1.*
> CC BY 4.0. <https://asr.iitm.ac.in/SPRING_INX>

How this was built — boundary extraction, the labeller bake-off, and every
experiment including the ones that failed:
<https://github.com/santhosh-005/tamil-eot>

The label tables ship in that repo under `labels/`, so the labeller bake-off
above reproduces from a clone with no audio and no GPU.