Converted from
ayousanz/cosy-voice3-onnx (Apache-2.0),
itself exported from
FunAudioLLM/Fun-CosyVoice3-0.5B-2512 (Apache-2.0).
Conversions: LLM backbones int4 (MatMulNBits, RTN block 32);
all other fp16 graphs upcast to fp32 (the fp16 graphs produce NaN
on CUDA and on onnxruntime >= 1.24 CPU).
Voice classic-zh is the official CosyVoice zero-shot prompt clip;
classic-ja is a fully synthetic clip generated with our own TTS;
sarah ships with the upstream export.
