Study On Objective Evaluation Of Chinese Speech Quality In The Presence Of Background Noise

Modern communication equipment may be exposed to various noise environments,and it is particularly important to effectively quantify the noise suppression effect of the terminal.In response to this problem,ITU-T Recommendation P.835 proposes a subjective test method that evaluates from three dimensions: S-MOS for speech quality,N-MOS for residual noise quality,and G-MOS for overall noise suppression performance.In terms of objective evaluation,there are three evaluation methods,ETSI EG 202 396-3,ETSI TS103 106 and ETSI TS 103 281,but their applicability to Chinese speech is still to be considered.Therefore,the purpose of this paper is to analyze the performance of each objective evaluation method for Chinese speech in the presence of background noise.This paper starts from the following four aspects:First,it is divided into two parts sample database and objective evaluation model of the three objective evaluation methods to sort out.The application scenarios of each model are analyzed,and the algorithm differences are analyzed.Secondly,the experimental environment is set up to solve the background noise simulation and communication network simulation problems in the laboratory scene,and provide the sampling basis for the evaluation of Chinese speech quality under background noise.Third,build a sample database based on Chinese speech.Firstly,the appropriate Chinese source is selected,and the Chinese speech samples transmitted by the terminal are recorded under various test conditions to obtain differentiated objective test results.Secondly,a reasonable subjective experimental program is designed to ensure that the results of subjective evaluation are effective and reliable.In the end,32 public audiences participated in subjective experiments,collecting a total of 13,824 sub-items,covering 48 test conditions,each of which obtained 96 MOS voting results in three dimensions.Fourthly,the objective test results are compensated by the mapping function of the curve fitting.The uncertainty of the subjective test results under each test condition is considered,and the performance of each objective model is evaluated by using statistical indicators.The experimental results show that ETSI TS 103 281 Model A has the best accuracy and correlation for the objective evaluation of Chinese speech quality in the presence of background noise.