Method for automatic evaluation of the voice signals quality with low bit rate voice coding

Abstract

Evaluation of the VSQ is an important problem. In known methods for еvaluation the quality of analog or digital voice signals with pulse-code modulation (PCM) are used. When automatic quality assessing, it is assumed that the parameters of voice signal are determined. Then the corresponding subjective assessment of VSQ is selected. Objective methods for quality assessing are obtained by comparing the reference signal and the signal transmitted through the communication channel. To eliminate the identified contradictions, it is proposed to use an objective method that provides the assessing of VSQ in the absence of a reference signal and without conversion to the PCM format, i.e. based on the analysis of the LBR voice digital stream. The analyzed digital stream is considered as a system of random variables, which is characterized by a set of mean values and a covariance matrix that make up the pattern of input implementation. The reference pattern also includes a set of mean values and a covariance matrix. The initial data for the formation of a reference pattern are digital streams formed according to a given protocol of LBR voice coding, having the highest VSQ. The proposed method is based on calculating the divergence between the reference pattern and input implementation pattern. At the training stage, several distorted samples with a known number of bit errors are formed from the reference sample. When comparing their patterns, the divergence values are calculated, which are put in compliance with the VSQ assessment. Two approaches have been identified for making the compliance: experimental and experimental-analytical. The choice of approach is determined by the presence of initial data characterizing the VSQ for particular vocoder. Then generated compliance is interpolated. At the assessment stage for each implementation is calculated the divergence between the explored pattern and the reference pattern, which is formed from the training sample. Based on the divergence, the VSQ assessment is calculated. To test the proposed method a reference sample of digital streams obtained by compressing voice signals with a LPC-10-2400 vocoder was formed. Based on the initial data, which are presented in Recommendation-R F.1112-1, the compliance between the values of sound intelligibility and divergence was compiled by experimental-analytical approach. The obtained correspondence was interpolated by the least squares method using a third degree polynomial.