---
language:
- lg
- ach
- nyn
- en
- mas
- sog
tags:
- text-to-speech
- xtts
- coqui
- uganda
license: other
---

# XTTS Ugandan Languages (xtts-ug-lang)

This model is a fine-tuned version of [Coqui's XTTS-v2](https://huggingface.co/coqui/XTTS-v2) designed to support multiple Ugandan languages. 

## Supported Languages
The model was trained with special language tokens natively injected into the vocabulary:
- **Luganda** (`[lg]`)
- **Acholi** (`[ach]`)
- **Masaaba** (`[mas]`)
- **Runyankore** (`[nyn]`)
- **Soga** (`[sog]`)
- **Ugandan English** (`[en-ug]`)

## Dataset
Trained on the `reuben256/votex-v1` dataset comprising over 248 hours of high-quality speech data.

## Training Details
- Fine-tuned using the XTTS GPT trainer.
- The base model's embedding matrices were successfully patched and resized to support the new vocabulary IDs.

## Usage
To use this with the standard Coqui `TTS` library, download the repository and load it locally. Make sure you use the patched `config.json` and `vocab.json` included in this repository. 

Because the underlying library has strict language validation, you should set `language="en"` in the API call but include your specific language token (e.g., `[lg]`) directly inside the text string.

### Python Example

```python
from TTS.api import TTS

# 1. Download the repo and set the paths to the local directory
model_dir = "./xtts-ug-lang"
config_path = f"{model_dir}/config.json"

# 2. Initialize the TTS API
tts = TTS(
    model_path=model_dir,
    config_path=config_path
)

# 3. Generate Audio
# Pass language="en" to bypass strict checks, but prepend the real language tag to the text.
text_to_speak = "[lg] Ekiseera kino kyamagero."

tts.tts_to_file(
    text=text_to_speak,
    speaker_wav="path/to/reference/speaker.wav", # Your reference audio for voice cloning
    language="en",
    file_path="output.wav"
)
print("Audio generated successfully!")
```