README.md

---
configs:
- config_name: default
  data_files:
  - split: test
    path: data/test-*.parquet
---

# Turn Eval

This dataset contains paired wake-word and turn audio for turn-level evaluation.
It has one `test` split and uses the default dataset configuration.

## Load the dataset

```python
from datasets import load_dataset

dataset = load_dataset("luvox-ai/turn-eval", split="test")
```

## Row structure

Each row represents one indexed wake-word/turn audio pair for one target voice.
The same `voice_id` can appear in multiple rows. For illustration, an anonymous
voice identifier could look like `speaker_001`; actual identifiers are dataset
values and are intentionally not listed in this card.

| Field | Type | Meaning |
|---|---|---|
| `voice_id` | `string` | Identifier of the target voice for this row. |
| `index` | `int32` | Pair index within the target voice. |
| `ww_audio` | `Audio` | Wake-word audio for the pair. |
| `turn_audio` | `Audio` | Turn audio used for speech-region annotation. |
| `segments` | `list<struct>` | Speech regions inside `turn_audio`. |

Each item in `segments` has this structure:

| Field | Type | Meaning |
|---|---|---|
| `start_ms` | `int32` | Inclusive region start, in milliseconds from the beginning of `turn_audio`. |
| `end_ms` | `int32` | Exclusive region end, in milliseconds from the beginning of `turn_audio`. |
| `label` | `string` | Speaker relationship: `target`, `non-target`, or `ambiguous`. |

Segments are ordered by start time and do not overlap. An empty `segments` list
means the turn was reviewed and no speech region was labeled. Multiple regions
remain in one row; the audio is not duplicated per segment.

## Label meanings

- `target`: speech attributed to the target voice identified by `voice_id`.
- `non-target`: speech attributed to another voice.
- `ambiguous`: speech whose target/non-target attribution is uncertain.

This card intentionally omits speaker identities, actual identifier values, and
per-voice statistics.