deep-dereverb

Deep learning based speech dereverberation

Final project to achieve the Sound Engineer degree from UNTREF (Universidad Nacional de Tres de Febrero).

In this work deep learning based speech dereverberation is studied. A type of convolutional neural network called autoencoder with jump connections is implemented, combining techniques of the current state of the art.

This autoencoder is trained to receive a magnitude spectrogram of an reverberated audio signal, and generate an amplitude mask which when is applyed over the initial spectrogram, produces the magnitude spectrogram of the dereverberated audio signal.

Input Spectrogram	Estimated Mask	Dereverberated Spectrogram	Target Spectrogram