On the Evaluation of the Conversational Speech Quality in Telecommunications

Abstract

We propose an objective method to assess speech quality in the conversational context by taking into account the talking and listening speech qualities and the impact of delay. This approach is applied to the results of four subjective tests on the effects of echo, delay, packet loss, and noise. The dataset is divided into training and validation sets. For the training set, a multiple linear regression is applied to determine a relationship between conversational, talking, and listening speech qualities and the delay value. The multiple linear regression leads to an accurate estimation of the conversational scores with high correlation and low error between subjective and estimated scores, both on the training and validation sets. In addition, a validation is performed on the data of a subjective test found in the literature which confirms the reliability of the regression. The relationship is then applied to an objective level by replacing talking and listening subjective scores with talking and listening objective scores provided by existing objective models, fed by speech signals recorded during the subjective tests. The conversational model achieves high performance as revealed by comparison with the test results and with the existing standard methodology “E-model,” presented in the ITU-T (International Telecommunication Union) Recommendation G.107.