How would you train for BW extension?

[kelseyjd]

I'm interested in training to convert 24 kHz mel spectrograms to 48 kHz waveforms (like HIFI-GAN2). Might not work without changing the architecture, but that's ok. How would you modify the config files to do this? I've already run the recipe through stage 1 to extract features with downsampled VCTK. Now I'm hesitating on how to modify the generator parameters to produce 2x length waveform with the HIFI gan config

[kan-bayashi]

You can simply increase upsample scale here.

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

Lines 38 to 39 in ffaa99f

	upsample_scales: [8, 8, 2, 2] # Upsampling scales.
	upsample_kernel_sizes: [16, 16, 4, 4] # Kernel size for upsampling layers.

E.g.,

 upsample_scales: [8, 8, 4, 2]         # Upsampling scales. 
 upsample_kernel_sizes: [16, 16, 8, 4] # Kernel size for upsampling layers.

[kelseyjd]

I see, thank you very much!

…

On Fri, Mar 3, 2023 at 7:17 PM Tomoki Hayashi ***@***.***> wrote: You can simply increase upsample scale here. https://github.com/kan-bayashi/ParallelWaveGAN/blob/ffaa99fe77d3b0703e5857177fd9b2ecc18cb0bd/egs/ljspeech/voc1/conf/hifigan.v1.yaml#L38-L39 E.g., upsample_scales: [8, 8, 4, 2] # Upsampling scales. upsample_kernel_sizes: [16, 16, 8, 4] # Kernel size for upsampling layers. — Reply to this email directly, view it on GitHub <#397 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ANZ6JON2SSKSBVATQSKYDHTW2KQ27ANCNFSM6AAAAAAVMWYTAQ> . You are receiving this because you authored the thread.Message ID: ***@***.***>