IDEAS AND TODOS

• mask multiply change to other operation??
• validate pydub on node
• rewrite perceptual quality to only use size and wetness

General Ideas

• Use formant changes as augmentation (are datasets biased towards male voices?)
• Train with Parameter Reverb
• Try CyclicLR Optimizer (https://docs.pytorch.org/docs/main/generated/torch.optim.lr_scheduler.CyclicLR.html#torch.optim.lr_scheduler.CyclicLR)

Discriminator Netwerk based on PEAQ. Use it as loss function (https://discuss.pytorch.org/t/train-a-neural-network-and-using-it-as-a-loss-function/156601/9)

• annotate data with attributes used for PEAQ?
• Use PESQ as additional label (is speech more important than music for us?)

from asteroid.models import ConvTasNet
import soundfile as sf
import numpy as np
import torch

input, sr = sf.read("mix.wav", dtype="float32")
input = torch.from_numpy(input).unsqueeze(0).requires_grad_()

print(input.grad)

model = ConvTasNet.from_pretrained("JorisCos/ConvTasNet_Libri1Mix_enhsingle_16k")

output = model(input)
output.backward(torch.ones_like(output))
print(input.grad)