change the feature extraction setting of vctk config

[bondio77]

Hi, first of all, thank you so much for providing pre-trained models through many experiments. But what I want to ask is, I want to fine-tune the pre-trained VCTK model with my multi-speaker dataset. In the VCTK config file, fft_size = 2048, hop_length = 300, win_length = 1024, but the config of the TTS model I trained is 1024, 256, 1024. When fine-tuning, will it work if I change the config file to 1024, 256, 1024 to match my TTS model? The sampling rate is 24000. Thank you!

[kan-bayashi]

Sorry for the late reply.
I think you should train from scratch for the following reasons:

• The difference of fft size or window size might be OK if you finetune the model
• The difference of hop length is critical since the hop length determines the upsampling layer structure.
• If you want to change hop length, you need to change the upsample layers as well.

Example:
hop length = 300 -> 5 * 5 * 4 * 3

ParallelWaveGAN/egs/vctk/voc1/conf/hifigan.v1.yaml

Line 40 in 8674037

upsample_scales: [5, 5, 4, 3] # Upsampling scales.

[bondio77]

thank you so much for answering me
i will try your recommend. thank you!