A Self-Supervised Convolutional Neural Network Approach for Speech Enhancement

Abstract

Enhancement of speech means modification to the

speech which is degraded by noise. Speech enhancement leads

to improvement in the intelligibility of speech to

human listeners. Deep learning techniques have drawn

tremendous attention for speech enhancement in recent years

which require clean speech along with noisy speech for

training purpose. However, availability of clean speech

signal in naturalistic scenarios is challenging. To ameliorate

it, this study proposes a deep neural network- based on speech

enhancement approach without the requirement of clean

speech to train the model called self-supervised learning. In

the proposed framework, two CNN-based speech enhancement

models have been deployed for two noisy conditions (babble

noise and machinery noise). This work has been

accomplished on two different datasets: IEEE speech corpus

distorted with real-time noise and recorded speech signals

in naturalistic environment. Experimental result demonstrates

that the proposed framework achieved significant

improvement in both subjective and objective measures.