monolithdm/kokoro-82m-kora

Kokoro-82M (ONNX) with per-phoneme durations

The ONNX export of Kokoro-82M with one extra output, pred_dur: the predicted duration of every input phoneme token, used to drive lip-sync for an animated avatar.

• Audio is unchanged. The graft adds a single Identity node exposing an existing integer tensor (/encoder/Cast_output_0) as a graph output; the waveform path is untouched. Verified bit-identical against the source export (59,400 samples, max difference 0).
• Units: one duration unit = 600 samples at 24 kHz = 25 ms. pred_dur has one value per input token (phoneme characters plus the start/end pad tokens).
• Files: onnx/model.onnx (fp32, WebGPU) and onnx/model_quantized.onnx (q8, CPU/WASM), plus the original config and tokenizer. Voices load from the upstream repo.

Credits

• Model: hexgrad/Kokoro-82M (Apache-2.0)
• ONNX export: onnx-community/Kokoro-82M-v1.0-ONNX (Apache-2.0)
• Graft method: scripts/graft_pred_dur.py from lorefoundry/tts-pack (Apache-2.0), run unmodified on the onnx-community export.