Design And Implementation Of Speech Codec Recognition Algorithm

Speech codec plays an important role in voice based communication systems.Therefore the techniques of speech codec recongnition have a lot of applications in the area of analysis of communication systems,recognition and terminals,counterwork of secure communications.Since the speech codec is designed with comprehensive consideration of the compression level,speech quality and complexity,different codec compression algorithms are different and have different effects on speech,these effects can be revealed in the encoded bitstream data or in the reconstructed speech after codec.Therefore,this paper designs and implements a recognition algorithm based on received bit stream and a recognition algorithm based on reconstructed speech.Speech codec recognition algorithm based on received bit streams is studied in the thesis.In the selection phase of feature parameters,some features that are often used in the encoding side recognition algorithm and can be calculated using existing tools are counted.Divided into three groups,the first group includes the autocorrelation coefficient,the center second moment,the Fourier transform,the binary ratio,the segment length is 20 kbits.The second group includes the mean,the autocorrelation coefficient,the Two,three,four order center distance,the binary ratio,the segment length is 8 kbits.The third group includes the mean,the variance,the first 20 coefficients of autocorrelation,the entropy,kurtosis,skewness,and the mean,variance,and skewness of the four subbands in the frequency domain.The segment length is 1 KB.Using the sequential floating forward search method,experiments were performed on different combinations of features in the three groups.By analyzing the experimental results,the optimal feature parameter combinations were finally determined.Algorithm implementation stage,the coded bit stream is preprocessed to form byte segment and bit segment,and the length of the segment is 20 kbits.Then the second third fourth order center distance,mean value,variance,Fourier transform value of the byte segment and binary ratio of the bit segment are extracted as the feature parameter,Using error correcting coding support vector machine as classifier.Use the TIMIT speech corpus as experimental data,the algorithm is simulated.The result of the experiment is that the GSM-HR,GSM-FR,GSM-EFR,AMR4.75,AMR7.95 and G.729 can be correctly identified as 100% correct.Because the compression algorithms of G.723.1₅3 and G.723.1₆3 are similar and the rates are relatively close,they are difficult to distinguish when they are identified.Speech codec recognition algorithm based on reconstructed speech is studied as well in the thesis.In the feature parameter selection stage,different from the traditional method to select a class of feature parameters,the algorithm considers combining multiple features to overcome the adverse effects of a single feature on the recognition algorithm.Experiments were conducted in two groups.The first group performed experiments on all permutations and combinations of MFCC static characteristics,first-order difference,second-order difference,energy,first order energy difference,and second order energy difference.The second group conducted experiments on the MFCC static features,the mean,the autocorrelation coefficient,the second three fourth order center distance,the Fourier transform,the variance,and the amplitude histogram.Through the analysis of the experimental results,the best feature parameter combination was finally determined.The classifier selects the neural network.By changing the parameter setting of the network,eight sets of experiments are performed and the final network parameter setting is determined.Algorithm implementation stage,the reconstructed speech is preprocessed to form a frame signal with a frame length of 32 ms.Then the MFCC,second order difference,first order difference of frame energy,autocorrelation coefficient,Fourier transform and variance combination of frame signal are extracted as feature parameters.Using the three-layer BPNN as the classifier and TIMIT speech corpus as experimental data,the algorithm is implemented.The experimental results show that GSM-HR,GSM-FR,GSM-EFR can be correctly identified by 100 percent.The recognition accuracy of G.729 encoder is 82.5 percent.