kokoro-82m-onnx-webgpu-nostft

Kokoro-82M, fp32 ONNX cut before the iSTFT (for onnxruntime-web WebGPU)

The fp32 onnx/model.onnx from onnx-community/Kokoro-82M-v1.0-ONNX at revision 1939ad2a, with the graph cut at decoder.generator.conv_post. The weights are unchanged.

Why: on onnxruntime-web 1.27 WebGPU, the full graph's iSTFT tail (ConvTranspose/ScatterND) returns zeros for every sample past about 65,528 (2.7 s at 24 kHz). Callers then have to feed short phrases, and that audibly flattens the prosody.

Inputs: input_ids int64 [1, n], style float32 [1, 256], speed float32 [1] (same as upstream). Output: spec_phase float32 [1, 22, frames].

iSTFT to run in the caller (n_fft 20, hop 5; istft_basis.json = the 22x20 inverse basis):

mag = exp(x[0:11]); ph = sin(x[11:22]); c = concat(mag*cos(ph), mag*sin(ph))   # 22 x F
y = conv_transpose1d(c, basis[22,1,20], stride=5)                               # 5(F-1)+20
y[1:20] /= window_sum[1:20]        # as exported: only the first window is normalised
waveform = 4 * y[10:-10]

This matches the full graph's waveform to within 6e-7 (max abs, CPU). Kokoro-82M is by hexgrad, released under Apache-2.0.

int8/ (2026-10-07)

Smaller copies of the files above for first-time downloads. Weights are stored as int8, symmetric, one scale per output channel, and each is followed by a DequantizeLinear node that turns it back into the original type (fp16 graphs dequantize straight to fp16, so their opset goes from 17 to 19). The graph, inputs and outputs are otherwise unchanged, and the maths runs at the original precision. Made with wq.py (skillsafe-app dev/local-character/slim/).

Measured on onnxruntime-web 1.27 WebGPU (Apple GPU):

• Kokoro body: 325 → 83 MB. 18 clips (6 sentences × 3 voices): UTMOS 4.385 vs 4.387 for the original, Whisper base.en WER 5.8% vs 4.2% (the gap is mishearing noise, mostly on the line both versions get wrong). Same speed.
• DreamShaper 8 LCM: 2.16 → 1.03 GB in total (UNet 861 MB, text encoder 124 MB, VAE decoder 50 MB). 8 prompts with fixed seeds give the same compositions and faces with small changes in detail (PSNR 20–31 dB against the original). 2.53 s per picture vs 2.08 s, because the dequantize runs on every UNet step.