NowLoadY/XRTranslate-OpenVoice-ONNX

XRTranslate OpenVoice ONNX

Community-maintained, reproducible ONNX conversions of official MyShell MeloTTS language checkpoints for XRTranslate's pure-Rust OpenVoice V2 provider. These are not official MyShell or NVIDIA artifacts.

Packages

Package	Language	Precision	Installed size	Base output
openvoice-v2-zh-onnx-fp16	Chinese (mixed English)	FP32 text / FP16 acoustic	460.9 MiB	44,100 Hz

Each package is self-contained: its language-specific frontend and feature graph, MeloTTS acoustic graph, OpenVoice V2 converter, reference speaker encoder, source tone embedding, licenses, and a machine-readable per-file SHA-256 manifest. XRTranslate returns converted audio at 22,050 Hz mono PCM16.

Reproducibility

The production graphs are exported in the existing torch211cu128 Conda environment with Python 3.10, PyTorch 2.11.0+cu128, CUDA 12.8, ONNX 1.22.0, ONNX Runtime 1.23.2, and opset 16. PyTorch is not upgraded or downgraded.

conda run -n torch211cu128 python third_party/openvoice_onnx_export/build.py `
  --language <language-key> `
  --output runtime/.temp/openvoice-onnx-export

Exact upstream commits, source hashes, graph tensor contracts, conversion versions, smoke-test results, and produced file hashes are recorded in each packages/<language-key>/v1/package-manifest.json.

Upstream and licenses

• MeloTTS and OpenVoice are MIT licensed.
• Each package manifest pins its exact MeloTTS checkpoint repository and commit.
• Source tone embeddings come from myshell-ai/OpenVoiceV2.
• Language-specific BERT/tokenizer, G2P, dictionary, and normalization notices are preserved inside that package.

See each package's licenses/ directory before redistribution.

Limitations

The packages require a compatible CUDA ONNX Runtime in XRTranslate. They are not standalone Python TTS APIs. Voice cloning transfers tone color; pronunciation, accent, and prosody remain controlled by the selected MeloTTS language pack.