TO-DO: Denoising approach - Githubissues

hyve9 / samplifi

MIT License

3 stars 0 forks source link

TO-DO: Denoising approach #3

Open hyve9 opened 5 months ago

hyve9 commented 5 months ago

Convolutional U-Nets can be used to predict a denoised output given the noisy input (https://www.sciencedirect.com/science/article/pii/S0952197622004973). The idea is

U-Net Model input: Noisy signal (here, from MBSG hearing loss attenuated signal)
Output for training: Original signal (may even pair with DEMUCS?)
Find the difference
Apply the difference to the original signal (assuming that the user's hearing loss will attenuate perceptually)

hyve9 commented 5 months ago

How should the diff be applied? Time-domain? Frequency-domain? MFCC? Something else?
What about both freq and time (hybrid approach)?