Project context and description:
Improving the quality of noise reduction models in the case of extremely low SNR (signal-to-noise ratio).
Main issues:
If the signal is highly noisy, the task of restoring the original signal becomes more difficult, since part of the signal is irretrievably damaged.
Solution:
We modified the models, used the DeepFeatureLoss approach and combined different losses. We used Gans to restore (generate) a signal lost due to noise.
Result:
Improved SDR and PESQ for cleaned files compared to noisy files.
Technology stack:
Pytorch, Demucs, WaveUNet, DCUNet, ConvTasNet, Hifi-GAN, UNet-GAN, etc.
Improving the quality of noise reduction models in the case of extremely low SNR (signal-to-noise ratio).
Main issues:
If the signal is highly noisy, the task of restoring the original signal becomes more difficult, since part of the signal is irretrievably damaged.
Solution:
We modified the models, used the DeepFeatureLoss approach and combined different losses. We used Gans to restore (generate) a signal lost due to noise.
Result:
Improved SDR and PESQ for cleaned files compared to noisy files.
Technology stack:
Pytorch, Demucs, WaveUNet, DCUNet, ConvTasNet, Hifi-GAN, UNet-GAN, etc.
