Why not use unvoiced information

[yangmingqi]

https://github.com/PlayVoice/NSF-BigVGAN/blob/d7204d8329e67e597f8856f3d0db596123a01d15/model/nsf.py#L305
The pitches extracted from the audio are all greater than zero, so the SineGen not use unvoiced information. Why do this, the unvoiced information is not important?

[MaxMax2016]

For train, pitch is from pyworld with uv;
For infer, pitch is from torchcrepe without uv，the uv of torchcrepe is not so accuracy to make some flaw;
Mel has the info of uv, so pitch without uv can be used for infer.