README.md

---
license: cc
datasets:
- speechcolab/gigaspeech
language:
- th
base_model:
- SWivid/F5-TTS
pipeline_tag: text-to-speech
tags:
- flow-matching
- f5-tts
- thai
- finetuning
---
<p align="center">
  <img src="assets/ThonburianTTSLogo.png" width="400"/><br>
  <img src="assets/looloo-logo.png" width="150" />
</p>


[🔊 Model Checkpoints](https://huggingface.co/biodatlab/ThonburianTTS) | [🤗 Gradio Demo](https://github.com/biodatlab/thonburian-tts/blob/main/gradio_app.py) | [📄 ThonburianTTS Paper](https://ieeexplore.ieee.org/document/11320472) | [Colab Notebook](https://colab.research.google.com/drive/1vIwNMjsyILluNT0l7I8KduS7S2Bhj9ra?usp=sharing) | [GitHub](https://github.com/biodatlab/thonburian-tts)

## **Thonburian TTS**

**Thonburian TTS** is a **Thai Text-to-Speech (TTS)** engine built on top of the [F5-TTS](https://github.com/SWivid/F5-TTS).  
It generates **natural and expressive Thai speech** by leveraging **Flow-Matching diffusion techniques** and can **mimic reference voices** from short audio samples. The system supports:

- **Thai language generation** (`language="th"`)
- **Reference-based voice cloning** using short audio clips
- High-quality synthesis with controllable speed and silence trimming


## **Model Checkpoints**

| Model Component        | Description                        | URL                                                                           |
| ---------------------- | ---------------------------------- | ----------------------------------------------------------------------------  |
| **F5-TTS Thai**        | Flow Matching-based Thai TTS models | [Link](https://huggingface.co/biodatlab/ThonburianTTS/tree/main/megaF5)      |
| **F5-TTS IPA**         | Flow Matching-based Thai-IPA TTS models | [Link](https://huggingface.co/biodatlab/ThonburianTTS/tree/main/megaIPA) |


## **Quick Usage**

### **Installation**

Install dependencies:

```bash
pip install torch cached-path librosa transformers f5-tts
sudo apt install ffmpeg
```

### **Clone GitHub**

```
git clone https://github.com/biodatlab/thonburian-tts.git
cd thonburian-tts
```

#### **Loading Thai Script based Models** 
```py
from flowtts.inference import FlowTTSPipeline, ModelConfig, AudioConfig
import torch

# Configure F5-TTS model
model_config = ModelConfig(
    language="th",
    model_type="F5",
    checkpoint="hf://biodatlab/ThonburianTTS/megaF5/mega_f5_last.safetensors",
    vocab_file="hf://biodatlab/ThonburianTTS/megaF5/mega_vocab.txt",
    vocoder="vocos",
    device="cuda" if torch.cuda.is_available() else "cpu"
)

# Basic audio settings
audio_config = AudioConfig(
    silence_threshold=-45,
    cfg_strength=2.5,
    speed=1.0
)

pipeline = FlowTTSPipeline(model_config, audio_config)
```


#### **Loading IPA based Models** 
```py
from flowtts.inference import FlowTTSPipeline, ModelConfig, AudioConfig
import torch

# Configure F5-TTS model
model_config = ModelConfig(
    model_type="F5",
    checkpoint="hf://biodatlab/ThonburianTTS/megaIPA/model_last_prune.safetensors",
    vocab_file="hf://biodatlab/ThonburianTTS/megaIPA/mega_vocab_ipa.txt",
    vocoder="vocos",
    device="cuda" if torch.cuda.is_available() else "cpu"
)

# Basic audio settings
audio_config = AudioConfig(
    silence_threshold=-45,
    cfg_strength=2.5,
    speed=1.0
)

pipeline = FlowTTSPipeline(model_config, audio_config)
```

## **Example Outputs**

<table>
  <tr>
    <td align="center">
      <a href="https://youtu.be/rvmNgh0-jws">
        <img src="https://img.youtube.com/vi/rvmNgh0-jws/0.jpg" width="320"><br>
        🎵 Sample 1 – Single-speaker Thai Normal Text
      </a>
    </td>
    <td align="center">
      <a href="https://youtu.be/jVz3EpRTn1U">
        <img src="https://img.youtube.com/vi/jVz3EpRTn1U/0.jpg" width="320"><br>
        🎵 Sample 2 – Single-Speaker Thai Code-mixed Text
      </a>
    </td>
    <td align="center">
      <a href="https://youtu.be/sbaOdMhz3Z4">
        <img src="https://img.youtube.com/vi/sbaOdMhz3Z4/0.jpg" width="320"><br>
        🎵 Sample 3 – Multi-Speaker Conversational Speech
      </a>
    </td>
  </tr>
</table>

---

## **Developers**

- [Looloo Technology](https://loolootech.com/)
- [Biomedical and Data Lab, Mahidol University](https://biodatlab.github.io/)

<p align="center">
  <img width="150px" src="assets/looloo-logo.png" />
</p>


## **Citation**

If you use **ThonburianTTS** in your research, please cite:

```
@INPROCEEDINGS{11320472,
  author={Aung, Thura and Sriwirote, Panyut and Thavornmongkol, Thanachot and Pipatsrisawat, Knot and Achakulvisut, Titipat and Aung, Zaw Htet},
  booktitle={2025 20th International Joint Symposium on Artificial Intelligence and Natural Language Processing (iSAI-NLP)}, 
  title={ThonburianTTS: Enhancing Neural Flow Matching Models for Authentic Thai Text-to-Speech}, 
  year={2025},
  volume={},
  number={},
  pages={1-6},
  keywords={Adaptation models;Codes;Accuracy;Error analysis;Phonetics;Robustness;Natural language processing;Text to speech;Noise measurement;Research and development;Thai text-to-speech;Flow matching;F5-TTS},
  doi={10.1109/iSAI-NLP66160.2025.11320472}}
```

```
Thura Aung, Panyut Sriwirote, Thanachot Thavornmongkol, Knot Pipatsrisawat, Titipat Achakulvisut, Zaw Htet Aung, "ThonburianTTS: Enhancing Neural Flow Matching Models for Authentic Thai Text-to-Speech", 2025 20th International Joint Symposium on Artificial Intelligence and Natural Language Processing (iSAI-NLP), Phuket, Thailand, 2025, pp. 1-6, doi: 10.1109/iSAI-NLP66160.2025.11320472.
```

## **License**

The **models** are released under the [Creative Commons Attribution Non-Commercial ShareAlike 4.0 License (CC BY-NC-SA 4.0)](LICENSE-CC-BY-NC-SA).

## Acknowledgement
We would like to acknowledge NSTDA Supercomputer Center (ThaiSC) project \#pv824003 for providing computing resources for this work.