The Study Of Improving Feature Extraction Algorithm In Speaker Recognition

With rapid development of information technology nowadays, a series of speech processing techniques and their applications have become the indispensable important component of information society. Speaker recognition, also called voiceprint recognition, is an important research topic in the field of speech processing technology. The speaker recognition technology belongs to biometrics, similar to the fingerprint recognition and iris recognition and so on. It has many advantages, including no need to remember, never forget, and easy to use, etc. Therefore it is widely used in judicial appraisal, medical applications, communications and other fields, making it considered as one of the most natural way of biometric recognition and identification.In order to achieve the purpose of determining the person that speaks, speaker recognition technology rather wants to extract the information which represents personal characteristics than pays more attention to the text symbols and semantic information contained in the speech signal. And for the feature extraction algorithm, its main task is to be able to select and research different effective and reliable feature vectors that can also distinguish the signal category. But so far there is not a simple and feasible method to separate the speaker’s personality characteristics from the speech signal completely. Thus to extract speaker’s personality characteristics so that different speakers can be distinguished better and solve the problem of low recognition rate under noise environment, the research in this article mainly focuses on researching extraction algorithm for feature parameters in speaker recognition parameter and its improvement.In this paper, first of all three basic characteristic parameters in the speaker recognition system, Mel frequency cepstral coefficient, linear prediction coefficient and the both mixing parameter named linear prediction Mel cepstral coefficient are applied to a complete platform which has built for speaker recognition. On the platform, the model for speaker recognition is hidden Markov model, which is commonly used in speech recognition for text-independent isolated words. For researching extraction algorithm in speaker recognition, firstly, put three basic feature parameters respectively in four different SNR (respectively15dB, lOdB,5dB, OdB) and no noise environments. Then use them as the basis for joining normalized short-time energy parameters as their auxiliary parameter, which contains some information about the speaker individuality. Finally complete the analysis of experimental results. The experimental results show that the recognition rate is not high, and in the case of low SNR recognition rate falls significantly.To extract speaker’s feature parameters which have optimal robust and degree of differentiation, two kinds of improved feature extraction algorithms have been presented in this paper. One is based on the Mel cepstrum composite coefficients and Fisher criterion with correlation distances, and the other one is based on the distributed discrete cosine transform and Fisher criterion with correlation distance too.Both of the two methods are based on the MFCC for optimization and improvement. The former method firstly put the normalized short-time energy parameters and the first order difference in MFCC to constitute composite vectors as new features, called MEL cepstrum composite coefficients. Then, in view of the high dimensional parameters, the algorithm for feature selection about Fisher criterion with correlation distance is introduced. The weighted algorithm is designed to lower dimension for Mel cepstrum composite coefficients. The latter improved algorithm uses distributed discrete cosine transform for MFCC extraction algorithm. To reduce its semantic information samples and enhance its parameter robustness under low SNR, the weighted algorithm about Fisher criterion with correlation distance is proposed again. The experimental results show that the two kinds of improved algorithms above increase the robustness of speaker recognition system and the recognition rate significantly. Therefore, the study of this article’s topic possesses certain practical significance in the speaker recognition.