Kokoro-82M v1.0 ONNX — mirror for the Reader PWA

A pinned copy of onnx-community/Kokoro-82M-v1.0-ONNX with one file changed, kept under this account so the app that loads it controls versioning. The app pins a commit of this repo; nothing here is meant to move underneath it.

file	what it is
onnx/model.onnx	fp32, patched for onnxruntime-web's WebGPU backend (see below)
onnx/model_uint8.onnx	official, unchanged — the CPU "Faster" model
onnx/model_quantized.onnx	official, unchanged — the CPU "Higher fidelity" model (q8)
config.json, tokenizer.json, tokenizer_config.json	official, unchanged
voices/*.bin	official, unchanged

The WebGPU patch

onnxruntime-web's WebGPU ConvTranspose returns wrong values for the vocoder's upsampling layers (/decoder/decoder/generator/ups.0 and ups.1): the output has the right length and is loud noise (peaks ~3e5–4e7 instead of ~0.6). Measured on AMD RDNA-3 with onnxruntime-web 1.22 and 1.30. Each of those ConvTranspose nodes is rewritten as the mathematically identical zero-insertion upsample followed by an ordinary Conv, which WebGPU computes correctly. CPU output is unchanged (cosine similarity 1.0, max difference 1.5e-6 per the original author).

The rewrite and its script come from DevAmarnadh/Kokoro-82M-v1.0-ONNX-webgpu (fix_kokoro_webgpu.py). The file here was rebuilt independently from the official onnx/model.onnx with that script and is byte-identical to the author's upload:

fix_kokoro_webgpu.py            sha256 aa24cf0c66b5afc08049b22fd627b8ca7557223cca027fd7b7f035835b8d0dbb
onnx/model.onnx (official)      sha256 8fbea51ea711f2af382e88c833d9e288c6dc82ce5e98421ea61c058ce21a34cb
onnx/model.onnx (this repo)     sha256 4c1aaa3cebf93f6cef797a36e978907305a3361288ba1c471400bddc0e0e2709

Rebuilt 2026-09-20 with onnx 1.23.0 / numpy 2.4.6 (Python 3.11).

Upstream

• Kokoro-82M: hexgrad/Kokoro-82M, Apache-2.0.
• ONNX export: onnx-community/Kokoro-82M-v1.0-ONNX.
• ORT bug tracking: microsoft/onnxruntime#29807.