DNN-based Source Enhancement to Increase Objective Sound Quality Assessment Score

TL;DR

This paper introduces a novel DNN training method that directly optimizes objective sound quality scores, like PESQ, using black-box optimization techniques to enhance perceived audio quality.

Contribution

It proposes a black-box optimization-based training scheme for DNNs that directly maximizes OSQA scores, overcoming the limitations of traditional MSE-based training.

Findings

OSQA scores significantly increased with the proposed method

The method improves sound quality without minimizing MSE

DNNs can be trained to directly optimize perceptual quality metrics

Abstract

We propose a training method for deep neural network (DNN)-based source enhancement to increase objective sound quality assessment (OSQA) scores such as the perceptual evaluation of speech quality (PESQ). In many conventional studies, DNNs have been used as a mapping function to estimate time-frequency masks and trained to minimize an analytically tractable objective function such as the mean squared error (MSE). Since OSQA scores have been used widely for sound-quality evaluation, constructing DNNs to increase OSQA scores would be better than using the minimum-MSE to create high-quality output signals. However, since most OSQA scores are not analytically tractable, \textit{i.e.}, they are black boxes, the gradient of the objective function cannot be calculated by simply applying back-propagation. To calculate the gradient of the OSQA-based objective function, we formulated a DNN optimization scheme on the basis of \textit{black-box optimization}, which is used for training a computer that plays a game. For a black-box-optimization scheme, we adopt the policy gradient method for calculating the gradient on the basis of a sampling algorithm. To simulate output signals using the sampling algorithm, DNNs are used to estimate the probability density function of the output signals that maximize OSQA scores. The OSQA scores are calculated from the simulated output signals, and the DNNs are trained to increase the probability of generating the simulated output signals that achieve high OSQA scores. Through several experiments, we found that OSQA scores significantly increased by applying the proposed method, even though the MSE was not minimized.