Kokoro-82M, split ONNX export for yomiscript
Kokoro-82M exported to ONNX in three parts, as used by the Kokoro Katarite of
yomiscript (@yomiscript/kokoro, the yomiscript CLI and the
@yomiscript/mcp server). The CLI and the MCP server download these files on first use.
| File | What it is |
|---|---|
models/enc.onnx, models/f0n.onnx, models/dec.onnx |
fp32: the text encoder and duration predictor, the F0 / energy predictor, the decoder. With the same inputs and noise they reproduce the original PyTorch Kokoro within its own run-to-run spread |
models/*_w16.onnx |
the same graphs with fp16 weights (163 MB in all), for the browser |
voices/*.bin |
the voice packs as raw float32 (510 x 1 x 256), identical to voices/ of onnx-community/Kokoro-82M-v1.0-ONNX |
The model is split so that a Katarite can replace the predicted F0 before the decoder runs. The STFT in the decoder is
an exact ONNX version of torch.stft / torch.istft (Kokoro's own ONNX-friendly STFT is an approximation). Export
scripts: demo/ys_web/export_split.py, exact_stft.py and quant.py in the yomiscript repository.
License: Apache 2.0, the license of Kokoro-82M (hexgrad). This repository only re-exports the weights.