This is onnx-community/Kokoro-82M-v1.0-ONNX with one change: the model's internal per-token duration tensor is exposed as a second graph output named pred_dur.

Kokoro is a StyleTTS2-family model: its duration head computes exactly how many audio frames each input token receives, and the vocoder renders precisely that. The Python kokoro package surfaces these durations as token timestamps; the stock ONNX export simply doesn't list the tensor as a graph output, so web runtimes (kokoro-js / transformers.js) never see it. Exposing it enables sample-accurate word timing (karaoke, RSVP, captions) with zero extra inference cost.

Pure graph surgery with the onnx package — no re-export, no re-quantization: the tensor /encoder/Clip_output_0 (the rounded, min-clamped per-token frame count that feeds the alignment CumSum) is aliased through a Cast-to-float32 node to a new graph output pred_dur, shape [1, num_input_tokens]. The Cast keeps the output float32 on every precision tier (the fp16 tiers carry it as float16 internally, which browsers can't reliably read); the values are small integer frame counts, so the cast is exact. The waveform output is bit-identical to the stock export (verified for every patched file).

Patched files: onnx/model.onnx (fp32), onnx/model_quantized.onnx (q8), onnx/model_q4.onnx, onnx/model_q4f16.onnx. The remaining variants (fp16, q8f16, uint8, uint8f16) are unmodified copies from the base repo.

Drop-in replacement — durations count frames of exactly 600 samples at 24 kHz, but you never need that constant: audio.length / sum(pred_dur) is the samples-per-frame at runtime.

Patch script: export_timestamps.py (in the Aloud repo, which uses this model for its neural karaoke).