Single Channel Speech Enhancement System using Convolutional Neural Network based Autoencoder for Noisy Environments

Abstract

Listen

Translate article

The proposed work involves forming a Convolutional Denoising Autoencoder (CDAE) for speech enhancement (SE). Compared to different information exchange methods, like through visible data, speech signals are more complicated. Noise in the environment causes a reduction in intelligibility and quality of speech signals during the conversation. Therefore, there is a need to enhance distorted speech signals. Over the years, many SE techniques have been developed to restore clean speech by reducing background noise. Denoising speech afflicted with noise is a complex process due to the variable characteristics of noise environments. Researchers also apply deep learning algorithms to enhance noisy signals. In this work, the architecture for CDAE is developed to enhance the noisy speech signal. A 2D input feature vector is fed to this architecture, which is obtained by estimating the magnitude spectrum of speech signals. The CDAE training set includes clean speech as the intended target and the magnitude of spectral values of a speech signal affected by noise as input. In this context, a performance metric is the perceptual evaluation of speech quality (PESQ). The suggested system gives higher PESQ values for considered noise environments than the existing approaches.