Automatic Speech Recognition-Driven Speech Enhancement Deep Neural Network: Automatic Speech Recognition-Driven Speech Enhancement Deep Neural Network
Translated title
Automatic Speech Recognition-Driven Speech Enhancement Deep Neural Network
Author
Term
4. term
Education
Publication year
2023
Submitted on
2023-06-02
Pages
57
Abstract
Speech Enhancement systems improve the quality and intelligibility of noisy speech signals. It has been proved that conventional loss functions such as MSE do not correlate highly with how humans perceive speech, and they do not perform well in subjective listening tests. On the contrary, objective metrics used as loss functions show better performance on objective tests. However, there is not always a correlation between the subjective and objective intelligibility tests. In this Thesis, an Automatic Speech Recognition (ASR) model is employed as a loss function in the Speech Enhancement system, aiming at closing the gap in intelligibility performance between subjective and objective tests. The hypothesis is that minimizing the Word Error Rate of a noisy speech signal will improve intelligibility in both cases.
Keywords
Documents