Hi @jaywalnut310

I've trained model with my datasets, I check files audio and labels that they are corrected and good.
But I trained model about 100k step about 7 days. Model misses some phone pronounce or quality phone not clearly as baby talks.

I not used dsp, add_blank. How about noise_scale params?

Almost step has same diff Spectro like image below:

Could you suggest me any advance to facing with it?

My words failed to express my gratitude for your helping!!