Kokoro-82M β€” ONNX mirror

ONNX exports of Kokoro-82M, mirrored for CDN stability as a dependency of the Omote SDK.

Provenance

Upstream chain (all Apache-2.0):

  1. hexgrad/Kokoro-82M β€” the original Kokoro-82M model (Apache-2.0).
  2. onnx-community/Kokoro-82M-v1.0-ONNX β€” community ONNX export of (1).
  3. This repo β€” mirror of ONNX artifacts derived from (2), hosted under omote-ai for CDN stability.

Files

File Size sha256 Description
kokoro_fp16.onnx 163,234,836 B 090dc6036fa19f587ce70782838ad41f226040abfdb6d961115fa66a404195cf fp16 single-file export β€” the artifact the Omote native runtime installs as kokoro/kokoro.onnx
kokoro_fp16_atan2guard.onnx 163,234,836 B 3574c63e84c2cad4eb3ba7a2641636ab0d9c4b64af8f5d1c4ed4e4547e3ac98a derived from kokoro_fp16.onnx (parent sha256 090dc603…): the vocoder's fp16 imag/real Div in its atan2 emulation gains a zero guard Where(Equal(real,0), 2^-14, real) β€” +2 nodes, same I/O, 0/40 corpus sentences non-finite at every thread count (the parent emits all-NaN audio on ~8/40 at ORT's default threads). Derivation script: scripts/kokoro-atan2-guard/patch_kokoro_atan2_guard.py in the Omote repo, deterministic. Apache-2.0 unchanged. Since 2026-09-13 the artifact the Omote native runtime installs as kokoro/kokoro.onnx
kokoro_int32.onnx 92,361,212 B e4c36c8a1c032cffcf15c5f7ee67be21c3ebd792717b7b6eab5ca339a4299e10 single-file web export (web smoke-harness production export)
kokoro_webgpu.onnx 1,149,264 B 35e95bb1cec095808bbfd8475f4f75e5e37948651c478be139a90d3c55b6ab12 WebGPU graph (ONNX external-data format β€” requires the .data file below alongside)
kokoro_webgpu.onnx.data 162,249,784 B ad8d64cee31dd04439f6d1844805310b5c736acd93ad8ec409a8e7d6e037d958 WebGPU external weight data

Voice data

Voice style tensors (e.g. af_heart) are not hosted in this repo. The Omote model registry sources voices/af_heart.bin directly from onnx-community/Kokoro-82M-v1.0-ONNX.

License

Apache-2.0, matching the upstream Kokoro-82M model and its ONNX export.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support