Method for Automatically Recognizing Speech based on Deep Neural Networks

Abstract

In order to increase system accuracy, this work offers the idea of voice enhancement using a hybrid DNN and a phase compensation approach. Initially, the aim of binary mask estimation was to locate the areas with SNR greater than 0 dB. Additionally, we have created an algorithm that can estimate the voice signal's active level even in the presence of significant noise. Convolutive or additive noise are also possible. In order to enhance the quality and understand ability of the boosted signal, noise reduction algorithms estimate the clean voice signal from the noisy recordings. In this work, it offers voice augmentation in various loud environments. The performance of the existing DNN approach is then compared in terms of WER and SNR.