thesis

thesis copied to clipboard

Add data augmentation

1

DDSP and NEWT papers don't mention it, but RAVE does: "We use dequantization, random crop and allpass filters with random coefficients as our data augmentation strategy."

vvolhejn

Conduct mean opinion score survey

1

Neural audio synthesis models are hard to compare automatically - a survey will be needed to show that the quality didn't decrease through our speedups

vvolhejn

TorchScript's performance seemed not to improve when forcing it to use both vCPUs, and DeepSparse [explicitly chooses not to use both](https://github.com/neuralmagic/deepsparse/issues/459). What if we change the number of CPUs?

vvolhejn

RAVE uses this, but the other models could too.

vvolhejn

vvolhejn

RAVE (and [Multiband MelGAN](https://arxiv.org/pdf/2005.05106.pdf) too) feeds the raw single-band waveform to the discriminator. Wouldn't it make sense to use multiband decomposition for the discriminator as well?

vvolhejn

Look into datasets

4

vvolhejn