Study Of Improving Feature Extraction Algorithm In Speaker Recognition

Speaker recognition is a broad of speech recognition. The basic idea is todetermine the identity of the speaker based on the speaker’s voicecharacteristics. In recent years, with advances in science and technology,requirement from all areas which is related to speaker recognition technologyis constantly improving. This makes the speaker recognition technology isfacing great challenges. On the one hand, the speaker recognition featureparameters will change with the person’s physical condition, emotionalcharacteristics and the environment when they are talking; On the other hand,the speaker recognition does not focus on semantic information, but paysattention to the speaker signal personality traits information. To accuratelyidentify the speaker’s identity, it must be accurately separated the semanticinformation and the speaker’s personality. But there is no one technology cancompletely separate them. In this paper, how to solve these problems werestudied.MFCC parameter can describe the spectral characteristics of the signalenvelope. But the signal spectral envelope is characterized mainly channelspeaker characteristics, ignoring the impact of the pitch frequencycharacteristic. To solve this problem, we put forward an improved algorithm,that is, when extracted MFCC parameter, we do not get the spectrum of the signal through Mel filter bank directly, but firstly we use the moving averagefilter for smoothing the signal spectrum to obtain a signal spectrum envelopeapproximated. The results obtained are then filtered by Mel filter. On this basis,we use multiple windows instead of Hamming window to calculate the changeof signal. Then we get a new feature parameter MTSMFCC. Experimentsshow that noise robustness and time robustness of the speaker recognitionsystem based on MTSMFCC are improved.In order to solve the problem that the recognition rate of singlecharacteristic parameter is low in noisy environments, this paper carried outthree aspects of fusion:1. In order to fully reflect the characteristic parameterof the dynamic characteristics of speech, we fusion the original MFCC and thefirst-order differential parameter MFCC. Then we get parameter Fusion1;2.Inorder to fully reflect the low-frequency information, frequency informationand high frequency information of voice, MFCC, IMFCC and MidMFCC werefused to obtain parameter Fusion2.3.On the basis of the previous two fusion,Fusion1and Fusion2were fused to get the new feature parameter NMFCC.The new parameter NMFCC not only fits auditory characteristics of the humanear, but contains the low-frequency information, frequency information andhigh frequency information of voice which is more fully reflect the speaker’spersonality. Experiments show that compared with Fusion1and Fusion2, therecognition rate of new feature parameter NMFCC there are different degreesof increase in noisy environments.