Need help for training a 44k model

[ILG2021]

I have trained a 44k model recently, but the inference result is a little strange, the audio seems raised the pitch.
I have change the mel parameters in ljspeech.yaml:
n_fft: 2048
n_feats: 128
sample_rate: 44100
hop_length: 512
win_length: 2048

matcha.yaml:
n_feats: 128

I use nvidia/bigvgan_v2_44khz_128band_512x as the vocoder.
@shivammehta25