Comparison of CNN-based Speech Dereverberation using Neural Vocoder

Abstract

Reverberation degrades the speech quality and intelligibility, particularly for hearing impaired people. In an automatic speech recognition (ASR) system, a dereverberation technique, which removes reverberation, is widely employed as a pre-processing to increase the performance of the ASR system. In this paper, we compare the performance of the CNN-based dereverberation method by applying various vocoders. The U-Net architecture is employed as the dereverberation technique. WaveGlow, MelGAN, and Griffin Lim are used as vocoders. Such vocoders play a role in converting speech features into speech samples in time domain, and are capable of generating high-quality speech from mel-spectrograms. In order to compare the results, PESQ was measured. As a result, it was confirmed that PESQ was higher than that of the reverberant speech when speech was synthesized with the reverberation removal and vocoder.