---
dataset_info:
  features:
  - name: audio
    dtype:
      audio:
        sampling_rate: 16000
  - name: text
    dtype: string
  - name: ref_index
    dtype: int64
  splits:
  - name: train
    num_bytes: 1000427885.04
    num_examples: 4730
  - name: validation
    num_bytes: 21703376.0
    num_examples: 179
  download_size: 1012364638
  dataset_size: 1022131261.04
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: validation
    path: data/validation-*
---
# Code-Quasar/Tn-TTS

Tunisian Derja speech for VoxCPM2 fine-tuning.

- `audio` — 16000 Hz mono, decoded by the `datasets` Audio feature
- `text` — transcript, every row prefixed with `(Tunisian Dialect)`
- `ref_index` — row index of a reference clip for in-context cloning, or -1

## Use

```bash
python data.py pull  # materializes wavs + manifests
```

Derived from corpora with their own licence terms; the voices belong to real speakers. Check before redistributing.