MeloTTS-ONNX

MeloTTS to ONNX Exporter

A script to convert MeloTTS models to ONNX format for use with sherpa-onnx.

Supported Languages

Note: English (en, en_v2, en_newest) and Chinese (zh) models include English words in their lexicon for mixed-language support. Monolingual models (ja, fr, es, ko) do not include English words.

Quick Start

# Install dependencies and export a model (default: ko)
./run.sh ko

# Export a specific language
./run.sh fr

Manual Usage

pip install -r requirements.txt
python -m unidic download  # required for Japanese

python export-onnx.py --language zh

Options

python export-onnx.py --language <LANG> [options]

  --language    zh | en | en_v2 | en_newest | ja | fr | es | ko  (default: ko)
  --output      Output directory  (default: ./onnx_exports/<LANG>/)
  --device      cpu | cuda  (default: auto-detect)
  --opset       ONNX opset version  (default: 18)

Output Files

onnx_exports/<lang>/
├── model.onnx      # ONNX model with embedded metadata
├── tokens.txt      # Phoneme symbol → index mapping
├── lexicon.txt     # Word → phoneme + tone mapping
└── metadata.json   # Model metadata (language_code, sample_rate, speakers, …)

lexicon.txt format

word  phone1 phone2 ...  tone1 tone2 ...

Chinese and English models include English words for mixed-language pronunciation. Monolingual models (ja, fr, es, ko) contain only target language vocabulary.

How It Works

The export script wraps the MeloTTS SynthesizerTrn model in a ModelWrapper that:

1. Accepts token IDs, tone information, speaker ID, and synthesis parameters
2. Sets BERT embeddings to zeros (C++ BERT inference is not available in sherpa-onnx)
3. Calls the internal VITS inference and exports via torch.onnx.export

Note: BERT embeddings are zeroed out, so voice quality may differ slightly from the original PyTorch model.

License

MeloTTS is licensed under MIT. See MeloTTS for details.