ThonburianTTS

<p align="center"> <img src="assets/ThonburianTTSLogo.png" width="400"/><br> <img src="assets/looloo-logo.png" width="150" /> </p>

🔊 Model Checkpoints | 🤗 Gradio Demo | 📄 ThonburianTTS Paper | Colab Notebook | GitHub

Thonburian TTS

Thonburian TTS is a Thai Text-to-Speech (TTS) engine built on top of the 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

Quick Usage

Installation

Install dependencies:

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

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

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
• Biomedical and Data Lab, Mahidol University

<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).

Acknowledgement

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