blurred validation output

[JohnHerry]

cards: 4 , batch size: 10 , lr: 1e-4
sample audio : 1s - 20s each, 16K audios, hop_size=200
mel-extractor: same as HiFiGAN
text tokenizer: phoneme tokens

I have trained to 250,000 steps but still can not get clear mel-spectrogram output. the validation output is blurred and diluteed artifacts,

has anybody run into such a situation? what is the reason? is this problem on training data preprocessing? or need some training parameter adjustment? wating for suggestions. thanks.