Samples for Audio Denoising with Deep Network Priors

A comparison of our unsupervised audio denoising based on deep network priors method with the classical unsupervised audio denoising baselines.

It seems that our results are cleaner and put less distortion to the speech signal compared to most unsupervised methods.

On the second table we present a comparison between our method to MMSE-LSA when the unvoiced parts at the edges of the clips are trimmed.

Our method seems indifferent to the change while MMSE-LSA suffers from voice distortions due the dependence on the noise statistics at the clip's beginning.

Audio Demo on Evaluation subset: (better to use headphones)

Audio Demo on Evaluation subset, clips with trimmed silence: (better to use headphones)