sbv2-speak-voices

sbv2-speak Voices

ONNX exports of Japanese Style-Bert-VITS2 text-to-speech voices, built to run with onnxruntime-web (or any other ONNX Runtime build) instead of requiring a Python backend.

Why ONNX?

Style-Bert-VITS2 voices are normally distributed as PyTorch checkpoints (.safetensors), which only run through a Python inference stack. Exporting to ONNX lets a voice run anywhere an ONNX Runtime is available — including entirely client-side, in a browser or Electron app, via onnxruntime-web, with no server or Python install needed. Each voice here was exported with torch.onnx.export + onnxsim (with a fix for a real onnxsim bug that otherwise silently corrupts audio content while still producing a valid-shaped file).

Usage

Each voice under voices/<id>/ has two files you need: model.onnx and style_vectors.npy.

1. Download both files for the voice you want.
2. Preprocess your input text into Style-Bert-VITS2's model inputs: phoneme IDs, tone IDs, a language-ID sequence, and Japanese BERT (ku-nlp/deberta-v2-large-japanese-char-wwm) features — the same preprocessing Style-Bert-VITS2 itself does before inference.
3. Load style_vectors.npy and select a style row (index 0 is "Neutral").
4. Run model.onnx in any ONNX Runtime build, passing those as inputs: x_tst, x_tst_lengths, sid (speaker ID, 0 for these single-speaker voices), tones, language, bert, style_vec, plus the generation knobs length_scale, sdp_ratio, noise_scale, noise_scale_w. This produces raw audio samples.
5. Write the samples out as audio (e.g. a WAV) at the model's sample rate.

For a complete, working implementation of steps 2-5 in JavaScript, see the sbv2-speak project's packages/inference package (runs in-browser via onnxruntime-web). For the reference PyTorch implementation these exports were traced from, see Style-Bert-VITS2 itself.

Source & license

Every voice here is a straight ONNX re-export (no retraining, no content changes) of a voice from itoolsjp/Style-Bert-VITS2-Voice-Assets — fully AI-generated synthetic voices, not based on any real person's voice. Per that repo's terms of use: personal use, commercial use, R-18/18G use, and embedding in other systems (which is what this repo does) are all permitted; selling the model itself, using it to impersonate a real person, defamation, and political/religious propaganda are not. See the source repo for the complete terms.