DevAmarnadh/Kokoro-82M-v1.0-ONNX-webgpu

Kokoro-82M v1.0 ONNX, patched for WebGPU

This is onnx-community/Kokoro-82M-v1.0-ONNX (onnx/model.onnx, fp32) with one graph rewrite so it produces correct audio on onnxruntime-web's WebGPU backend.

The bug

ORT WebGPU's ConvTranspose gives wrong values for the vocoder upsampling layer generator/ups.1 (kernel 12, stride 6). The output keeps the correct length and duration, but every sample is wrong: peaks around 3e5 instead of about 0.6, and the audio is unintelligible. This was measured on AMD RDNA-3 with onnxruntime-web 1.22 and 1.30, via transformers.js 3.7.5, 3.8.1 and 4.3.0.

The fix

Each vocoder ConvTranspose is replaced with an identical zero-insertion upsample followed by an ordinary Conv (flipped, transposed kernel). On CPU the output matches the original (cosine similarity 1.0, max difference 1.5e-6). See fix_kokoro_webgpu.py.

	peak	rms	Moonshine transcript WER
original, wasm	0.559	0.069	0.11
original, webgpu	309315	767	1.0 (empty)
patched, webgpu	0.555	0.068	0.11 (identical text)

fp16 and q4f16 are not fixed by this change: they still produce NaN on WebGPU, which is a separate problem.

Usage (kokoro-js)

const tts = await KokoroTTS.from_pretrained('DevAmarnadh/Kokoro-82M-v1.0-ONNX-webgpu', { device: 'webgpu', dtype: 'fp32' });

Voice files still load from onnx-community/Kokoro-82M-v1.0-ONNX, which kokoro-js hardcodes, so they're unchanged.