FastPitch VCTK 44k training

[lheuveline]

Hi,

I'm trying to train a 44k version of FastPitch on VCTK dataset. I'm using current VCTK recipe for FastPitch, with sample rate adjusted.
The training seems to go well, but my current number of steps is quite high with almost unintelligible speech (tested with griffin-lim and pretrained LJSpeech HifiGAN).

Here is my eval on last epoch (epoch = 544, global step = 1483450, batch_size = 16) :

--> EVAL PERFORMANCE
| > avg_loader_time: 0.02167 (+0.00049)
| > avg_loss_spec: 3.83201 (-0.00796)
| > avg_loss_dur: 0.00863 (-0.00009)
| > avg_loss_pitch: 0.23789 (-0.00053)
| > avg_loss_aligner: 2.42255 (-0.00234)
| > avg_loss_binary_alignment: 0.84863 (+0.00037)
| > avg_loss: 7.34971 (-0.01055)
| > avg_duration_error: 7.76402 (+0.00001)

I checked pretrained 22k FastPitch checkpoint and found that last step on pretrained model's checkpoint is 900000.
I think that increasing sample rate from 22k to 44k should probably increase the training time. But should I be worried about this results, or unintelligble speech at this step can be expected ?