VITS2 inference

I took the code from this repo to do inference without any additional files

Prerequisites

1. Python >= 3.10 (I used 3.11)
2. Clone this repository
3. Install espeak and requirements: pip install -r requirements.txt
   • Windows users: espeak installation can be tricky. See this issue for installation guide
4. Add your onnx model into folder, with your config file
5. Adjust your symbols in text/symbols.py if needed
6. For Russian accent support: pip install ruaccent

Usage

Command-line Arguments

python infer_onnx.py --model <model_path> --config-path <config_path> [options]

Required Arguments

• --model: Path to the ONNX model file
• --config-path: Path to the configuration JSON file

Optional Arguments

• --output-wav-path: Path to save the output WAV file (default: output/output.wav)
• --text: Text to synthesize (if not provided, interactive mode will be started)
• --lang: Language code (default: en)
• --sid: Speaker ID for multi-speaker models (default: None)
• --play: Play audio after synthesis (default: False)
• --accent: Apply Russian accent marks (only works with --lang ru)
• --accent-model: RUAccent model size (default: turbo2, options: tiny, tiny2, tiny2.1, turbo2, turbo3, turbo3.1, turbo, big_poetry)
• --accent-dictionary: Use dictionary for accents (default: True)
• --accent-device: Device to run ruaccent model on (default: CPU, options: CPU, CUDA)

Examples

1. Save only (default):

python infer_onnx.py --model model_name.onnx --config-path config.json --text "Всем привет!" --lang ru

2. Save and play:

python infer_onnx.py --model model_name.onnx --config-path config.json --text "Всем привет!" --lang ru --play

3. Interactive mode, save only:

python infer_onnx.py --model model_name.onnx --config-path config.json --lang ru

4. Interactive mode, save and play:

python infer_onnx.py --model model_name.onnx --config-path config.json --lang ru --play

5. With Russian accent marks:

python infer_onnx.py --model model_name.onnx --config-path config.json --text "Я говорю по-русски. Это замок на двери." --lang ru --accent

Russian Accent Support

For better Russian speech synthesis, this tool supports automatic accent placement using the ruaccent library. This is important because Russian word stress is not fixed and can significantly affect pronunciation.

To use this feature:

1. Install the ruaccent library: pip install ruaccent
2. Add the --accent flag when using Russian language (--lang ru)
3. Optionally customize the accent model with additional parameters:
   • --accent-model: Choose different model sizes (smaller = faster, larger = more accurate)
   • --accent-dictionary: Enable/disable dictionary usage
   • --accent-device: Run on CPU or CUDA (requires onnxruntime-gpu)

Notes

• The output directory will be created automatically if it doesn't exist
• In interactive mode, you can type 'quit' to exit the program
• Audio playback requires a working audio device on your system
• For GPU acceleration of ruaccent, install onnxruntime-gpu: pip install onnxruntime-gpu