---
language:
- vi
license: apache-2.0
tags:
- text-to-speech
- tts
- vietnamese
- voice-cloning
- audio
- speech-synthesis
library_name: transformers
base_model: Qwen/Qwen2-0.5B
datasets:
- custom
pipeline_tag: text-to-speech
---

# Vira-TTS

Vietnamese Text-to-Speech model with Voice Cloning capability, finetuned from [MiraTTS](https://huggingface.co/YatharthS/MiraTTS) for Vietnamese language.

## Model Description

Vira-TTS is a neural TTS model that can synthesize natural Vietnamese speech from text while cloning the voice characteristics from a reference audio sample.

| Property | Value |
|----------|-------|
| Base Architecture | Qwen2-0.5B |
| Audio Codec | Fash-BiCodec |
| Sample Rate | 16kHz (native), 48kHz (with FlashSR) |
| Language | Vietnamese |

## Features

- **Zero-shot Voice Cloning**: Clone any voice from a short reference audio (3-12 seconds recommended)
- **Vietnamese Optimized**: Finetuned specifically for Vietnamese pronunciation and prosody
- **Text Normalization**: Automatic conversion of numbers and abbreviations to spoken form
- **High Quality Output**: 48kHz audio with FlashSR upsampling

## Usage

### Installation

```bash
pip install git+https://github.com/iamdinhthuan/Vira-tts.git
```

### Quick Start

```python
from mira.model import MiraTTS

# Model will be downloaded automatically
mira_tts = MiraTTS('model_pretrained')

# Provide reference audio for voice cloning
reference_audio = "speaker.wav"
text = "Xin chào, đây là giọng nói tiếng Việt."

context_tokens = mira_tts.encode_audio(reference_audio)
audio = mira_tts.generate(text, context_tokens)

# Save output
import soundfile as sf
sf.write("output.wav", audio.float().cpu().numpy(), 48000)
```

### Web UI

```bash
python app.py
```

### CLI

```bash
python infer.py --text "Xin chào" --reference speaker.wav --output output.wav
```

## Text Normalization

The model includes automatic Vietnamese text normalization:

| Input | Output |
|-------|--------|
| `Năm 2024` | `Năm hai nghìn không trăm hai mươi tư` |
| `100.000 VNĐ` | `một trăm nghìn đồng` |
| `TP.HCM` | `thành phố Hồ Chí Minh` |

## Requirements

- Python >= 3.10
- CUDA compatible GPU (recommended: 6GB+ VRAM)
- Dependencies: lmdeploy, fastaudiosr, ncodec, gradio

## Limitations

- Best performance with reference audio 3-10 seconds long
- Reference audio should be clear without background noise
- Some rare Vietnamese words may not be pronounced correctly

## Citation

```bibtex
@misc{vira-tts,
  author = {Dinh Thuan},
  title = {Vira-TTS: Vietnamese Text-to-Speech with Voice Cloning},
  year = {2025},
  publisher = {HuggingFace},
  url = {https://huggingface.co/dolly-vn/Vira-TTS}
}
```

## Acknowledgements

- [MiraTTS](https://github.com/ysharma3501/MiraTTS) - Original model architecture
- [Spark-TTS](https://huggingface.co/SparkAudio/Spark-TTS-0.5B) - Base TTS model
- [FlashSR](https://github.com/ysharma3501/FlashSR) - Audio super-resolution
- [LMDeploy](https://github.com/InternLM/lmdeploy) - Inference optimization
- [soe-vinorm](https://github.com/vinhdq842/soe-vinorm) - Vietnamese text normalization

## License

MIT

## Contact

- GitHub: [@iamdinhthuan](https://github.com/iamdinhthuan)
- Repository: [Vira-TTS](https://github.com/iamdinhthuan/Vira-tts)