Research Of Speech Quality Assessment Based On Deep Learning

Speech is indispensable in people's information communication.The quality of speech directly determines whether the information communication is smooth.Therefore,how to effectively evaluate the quality of the output speech is the goal of researchers at home and abroad.Aiming at the problem that the existing nonintrusive speech quality evaluation has low correlation with the subjective evaluation result and cannot meet the market demand,this paper studies the non-intrusive speech quality evaluation.This paper researchs a speech quality evaluation method combining stack autoencoder(SAE)and back propagation(BP)neural network.The method is implemented by BP neural network and a deep neural network composed of SAE.The essential features of speech are extracted by the SAE and then the feature is mapped to the subjective MOS score through the BP neural network.The simulation results show that the proposed method reduces the mean square error and improves the correlation coefficient with subjective evaluation compared with the existing itu-t p.563 and speech quality evaluation method based on fuzzy directed graph support vector machine(FDGSVM),but increases the evaluation time.Aiming at the problem of the long evaluation time of the above methods,this paper adopts the elite selection and adaptive step-and-step improved firefly algorithm to integrate into BP neural network to solve the problem.And this paper proposes a new speech quality assessment method based on the combination of SAE and improved backpropagation.The essential features of speech are extracted by the SAE and this feature is mapped to the subjective MOS score by improved backpropagation.The simulation results show that compared with the speech quality evaluation method combined with SAE and BP,the evaluation time is reduced by 67.30%.In the speech quality evaluation of the in-set language,the correlation coefficient with subjective evaluation is increased by 16.57%.The mean square error was reduced by 10.06%,but in the speech quality evaluation of the out-set languages,the correlation coefficient with subjective evaluation deteriorated by 31.23%,and the mean square error deteriorated by 22.85%.This method is suitable for the speech quality evaluation of the in-set language,but not for the speech quality evaluation of out-set languages.