Speech Dereverberation Using U-Net and Minimum-Phase Inverse Prefiltering

Reverberation in enclosed spaces degrades audio quality, particularly in speech signals, due to the overlap of direct sound with reflected waves. This paper aims to investigate a method for improving U-Net efficiency by using minimum-phase prefiltering. The method is especially effective for environments with high RT<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">60</inf> values, and it can be applied in scenarios where the Room Impulse Response (RIR) has been estimated. It involves a U-Net architecture for non-linear filtering of the all-pass component of the RIR, using spectrograms as input and output. The proposed approach improves dereverberation performance and restores the clarity of audio signals.