Technology And System Development Of Voice Quality Objective Evaluation

With the rapid development of communication technology,voice system is widely used in people's daily life.It occupies a very important position in people's daily information exchanged.Voice quality is the key to determine the user experience and the important standard to evaluate the performance of a communication system.So the voice quality must be accuratly and flexibly evaluated.Based on the speech enhancement technology,the pattern recognition technology and the objective quality evaluation technology of the reference source,a non-reference source speech quality evaluation system which can adapt to the ambient noise is built under the MFC architecture.It can be used for pure background noise and pure voice detection.The experimental results show that the correlation between subjective and objective scores is 56.9% higher than that of P.563 algorithm in shortwave communication environment.The main work and contributions of this paper are as follows:1.Using the MCRA algorithm and spectral subtraction algorithm to deal with the noisy audio and combined with the PESQ objective evaluation algorithm of the reference source,an objective evaluation algorithm of the non-reference source is obtained.Experiments show that the correlation between subjective and objective scores is 38.1% higher than that of P.563,and it is found that it can not deal with pure background noise audio,and the objective score of pure speech is low.2.Considering the similarity between Gaussian functions in GMM model,a GMM model reconstruction algorithm is proposed.By using the neighborhood concept proposed in this paper,the similar Gaussian function in the two models is eliminated,and then the GMM is reconstructed.The reconstructed GMM model is more discriminative,and the computational complexity is reduced due to the decrease of the Gaussian mixture.3.A GMMSVM classification method for semi-supervised learning for background noise,noisy speech and pure speech classification is proposed.The method uses background noise and pure voice GMM reconstruction model to describe the three types of audio,and provides input characteristics to SVM.Through the iterative continuous model learning until the preset number of iterations,the method finally gets the updated background noise and pure voice GMM reconstruction model and SVM three classification model.The algorithm can adapt to the environmental noise by a small amount of artificial annotation,and get the better classification effect.Experiments show that the classification accuracy is higher than that of GMM,SVM and GMMSVM in four scenarios by 20.8% ~ 35.1%.4.The semi-supervised learning GMMSVM classifier is combined with the time domain rules and the estimated SNR rules to perform pure background noise,noisy speech and pure speech pre-detection for the evaluation audio,and then it is combined with the objective evaluation of the non-reference source in 1 to obtain an objective evaluation algorithm of non-reference source.Experiments show that under the shortwave communication system,the subjective and objective score correlation ratio of the algorithm is 14.7% higher than that of the algorithm in 1,and is 56.9% higher than that of P.563.5.Based on the algorithm in 4,this paper builds an objective evaluation system without reference source under MFC architecture.As a result of multi-file processing,the system uses a multi-threaded technology,which can improve the system computing speed.