Vietnamese G2P is handled by vig2p.

This repository includes both the lightweight inference package and the Vietnamese fine-tuning recipe. Training instructions are in TRAINING.md.

git clone https://github.com/iamdinhthuan/Kokoro-Vietnamese.git
cd Kokoro-Vietnamese
pip install -e .

For CUDA, install the PyTorch build that matches your machine first.

For training utilities:

pip install -e ".[training]"

The CLI downloads these files from contextboxai/Kokoro-Vietnamese when local paths are not provided:

• kokoro_vi.pth
• kokoro_vi.onnx
• kokoro_vi_voicepack.pt
• config.json

Use these names with voice=... in Python or --voice ... in the CLI.

import soundfile as sf

from kokoro_vietnamese import KokoroVietnamese

tts = KokoroVietnamese(device="cuda", voice="diem_trinh")

audio, phonemes = tts.synthesize(
    "Giữa một buổi chiều yên tĩnh, cô ấy kể lại câu chuyện bằng một giọng nói ấm áp và chậm rãi."
)

sf.write("outputs/sample.wav", audio, 24000)
print(phonemes)

The package uses the same Kokoro runtime path as the original Gradio predictor. Pass normalize_peak=0.95 only if you want optional peak limiting before writing WAV files.

audio, phonemes = tts.synthesize("Tường nhà khách.", normalize_peak=0.95)

Use another voice:

tts = KokoroVietnamese(device="cuda", voice="mai_linh")
audio, phonemes = tts.synthesize("Hôm nay trời trong xanh, gió thổi nhẹ qua hiên nhà.")

Use local files instead of downloading from Hugging Face:

tts = KokoroVietnamese(
    device="cuda",
    model_path="kokoro_vi.pth",
    voicepack_path="voicepacks/diem_trinh.pt",
    config_path="config.json",
)

kokoro-vietnamese \
  --text "Giữa một buổi chiều yên tĩnh, cô ấy kể lại câu chuyện bằng một giọng nói ấm áp và chậm rãi." \
  --output outputs/sample.wav \
  --voice diem_trinh \
  --device cuda

kokoro-vietnamese --list-voices

kokoro-vietnamese \
  --batch-file texts.txt \
  --voice diem_trinh \
  --output-dir outputs/batch

Install ONNX dependencies:

pip install -e ".[onnx]"

Export the PyTorch checkpoint to ONNX:

kokoro-vietnamese-export-onnx \
  --output outputs/kokoro_vi.onnx

Run ONNX Runtime inference:

kokoro-vietnamese-onnx \
  --text "Tường nhà khách đã được sơn lại." \
  --onnx outputs/kokoro_vi.onnx \
  --output outputs/onnx.wav \
  --device cpu \
  --print-phonemes

When --onnx is omitted, the command downloads kokoro_vi.onnx from the Hugging Face model repo. Install onnxruntime-gpu and pass --device cuda to use CUDAExecutionProvider when available.

kokoro-vietnamese-gradio --device cuda --share

If AlbertModel fails to import in an old environment:

pip install -U git+https://github.com/iamdinhthuan/Kokoro-Vietnamese.git "transformers>=4.48,<5"