README.md

---
language: [ar]
license: apache-2.0
base_model: openbmb/VoxCPM2
pipeline_tag: text-to-speech
tags: [tunisian, derja, arabic-dialect, voxcpm, tts]
---

# Code-Quasar/voxcpm-tn

VoxCPM2 fine-tuned for **Tunisian Derja**.

- base: `openbmb/VoxCPM2` (2B, tokenizer-free, 48 kHz)
- method: full fine-tuning, 147 steps
- data: ~? h Tunisian read speech, 16 kHz mono

## Important: the dialect tag

Every training transcript was prefixed with `(Tunisian Dialect)`, so **untagged text is
out of distribution**. Always prefix it:

```python
from voxcpm import VoxCPM

model = VoxCPM.from_pretrained("Code-Quasar/voxcpm-tn", load_denoiser=False)
wav = model.generate(text="(Tunisian Dialect) عسلامة، شنوة أحوالك اليوم؟")

import soundfile as sf
sf.write("out.wav", wav, 48000)
```

On a GPU with under ~8 GB, disable compilation:

```python
import os
os.environ["TORCHDYNAMO_DISABLE"] = "1"
```

## Limitations

Trained on a small corpus of read speech, so expect limited prosodic range and
weaker long-form phrasing. Derived from source corpora with their own licence
terms; the voices belong to real speakers.