Introduction

About MeloTTS

MeloTTS is a high-quality, open-source text-to-speech system developed by MyShell AI. It is built on top of the VITS/VITS2 architecture and uses BERT-based linguistic features to produce natural-sounding speech. MeloTTS supports multiple languages and is designed to be fast enough for real-time CPU inference.

Strengths of the original MeloTTS:

• High naturalness and expressiveness in synthesized speech
• Fast inference — runs in real-time even on CPU
• Lightweight and easy to deploy
• Supports multiple languages (English, Chinese, Japanese, Korean, Spanish, French)
• Permissive MIT license, suitable for both commercial and non-commercial use

Limitations of the original MeloTTS:

• Not natively optimized for Vietnamese phonology (tones, phonemes)
• The default English/multilingual phonemizer does not handle Vietnamese tones and diacritics correctly
• No built-in support for Vietnamese-specific linguistic preprocessing

MeloTTS Vietnamese

MeloTTS Vietnamese is a version of MeloTTS specifically optimized for the Vietnamese language. It inherits the high-quality and fast-inference characteristics of the original model while introducing targeted improvements to handle the unique phonological properties of Vietnamese — including its 6 tones, complex vowel system, and syllable structure.

This model is designed to produce natural, accurate Vietnamese speech and can be easily fine-tuned on custom Vietnamese datasets.

Technical Features

• Uses underthesea for Vietnamese text segmentation
• Integrates PhoBERT (vinai/phobert-base-v2) to extract Vietnamese linguistic features
• Full support for Vietnamese language characteristics:
  • 45 symbols (phonemes)
  • 8 tones (7 tonal marks and 1 unmarked tone)
  • All defined in melo/text/symbols.py
• Text-to-phoneme conversion:
  • Based on the Text2PhonemeSequence library
  • An improved higher-performance version is available at Text2PhonemeFast

Fine-tuning from Base Model

This model was fine-tuned from the base MeloTTS model by:

• Replacing phonemes not found in English/Vietnamese with Vietnamese-specific phonemes
• Specifically replacing Korean phonemes with their corresponding Vietnamese equivalents
• Adjusting model parameters to match Vietnamese phonetic characteristics
• GitHub: MeloTTS Vietnamese

Training Data

• The model was trained on the Infore dataset, consisting of approximately 25 hours of speech
• Note on data quality: This dataset has several limitations including suboptimal voice quality, missing punctuation, and imprecise phonetic transcriptions. However, when trained on internal/private high-quality data, results are significantly better.

Downloading the Model

The pre-trained model can be downloaded from Hugging Face:

• MeloTTS Vietnamese on Hugging Face

---

Usage Guide

Part 1: Inference

1. Clone the Repository and Install Dependencies

git clone https://github.com/manhcuong02/MeloTTS_Vietnamese.git
cd MeloTTS_Vietnamese
pip install -r requirements.txt

2. Download the Pre-trained Model

Download the model checkpoint and config from Hugging Face and place them in your desired directory.

3. Run Inference

Refer to the notebook test_infer.ipynb for a full example. Basic usage:

from melo.api import TTS

# Speed is adjustable
speed = 1.0

# You can set device to 'cpu', 'cuda', 'cuda:0', or 'mps'
device = "cuda:0"  # Will automatically use GPU if available

# Load the Vietnamese TTS model
model = TTS(
    language="VI",
    device=device,
    config_path="/path/to/config.json",
    ckpt_path="/path/to/G_model.pth",
)
speaker_ids = model.hps.data.spk2id

# Convert text to speech
text = "Nhập văn bản tại đây"
output_path = "output.wav"
model.tts_to_file(text, speaker_ids["speaker_name"], output_path, speed=speed, quiet=True)

---

Part 2: Training & Fine-tuning

1. Data Preparation

The full data preparation process is detailed in docs/training.md. At minimum, you need:

• Audio files (recommended sample rate: 44100 Hz)
• A metadata file in the following format:

  path/to/audio_001.wav |<speaker_name>|<language_code>|<text_001>
  path/to/audio_002.wav |<speaker_name>|<language_code>|<text_002>
  

2. Data Preprocessing

Run the preprocessing script to prepare training data:

python melo/preprocess_text.py \
  --metadata /path/to/text_training.list \
  --config_path /path/to/config.json \
  --device cuda:0 \
  --val-per-spk 10 \
  --max-val-total 500

Alternatively, use the shell script melo/preprocess_text.sh with appropriate parameters.

3. Start Training

Follow the training instructions in docs/training.md.

---

Code & Fine-tuning

The Vietnamese adaptation, code implementation, and fine-tuning of this model were developed by Nguyễn Mạnh Cường.

• GitHub: manhcuong02
• Repository: MeloTTS Vietnamese

---

Audio Examples

Listen to sample outputs from the model:

Sample 1

"Buổi sáng ở thành phố bắt đầu bằng tiếng xe cộ nhộn nhịp và ánh nắng nhẹ xuyên qua những tòa nhà cao tầng."

Sample 2

"Người đi làm vội vã, học sinh ríu rít trò chuyện, còn quán cà phê góc phố thì thoang thoảng mùi thơm dễ chịu."

Sample 3

"Cuối cùng, hãy thử thì thầm một câu thật nhẹ nhàng, rồi bất ngờ chuyển sang giọng nói to, rõ và đầy năng lượng."

---

License

This project is licensed under the MIT License, consistent with the original MeloTTS project. It may be used for both commercial and non-commercial purposes.

Acknowledgements

This implementation is based on TTS, VITS, VITS2, and Bert-VITS2. We appreciate their outstanding work.