Research Of Speaker-Recognition Technology On Vector Quantization

Speaker recognition is the processing of automatically recognizing which is speaking by using speaker specific information included in speech signal. Speaker recognition has broad application foreground in many fields such as electric business and information security. In general, it can be classified into speaker identification and speaker verification according to decision modes. This thesis focuses attention on research of text-independent speaker recognition technology based on vector quantization.LBG algorithm is one of the common and important methods used in speaker recognition. But the main drawback of LBG algorithm is that it often gets trapped in local optima that are significantly worse than the global optimum. This thesis presents a randomized local search algorithm (RLS) for the vector quantization. Results indicate that the proposed algorithm is easy to implement and competitive with the best clustering methods currently. In addition, it is demonstrated to be more effective in the clustering for speech parameters and be obtained better codebook quality in comparison with LBG algorithm. The proposed algorithm in this thesis also shows a new idea in designing the best codebook for solving more complex problem in speaker recognition.Two speech corpuses which include text-independent speech recorded from twenty male speakers and ten female speakers respectively were built on experimentation. Some factors such as speed, volume and time interval which affect the performance of speaker identification system were taken into consideration.In addition, the properties and extraction methods of some common feature parameters are studied in detail. In particular, two kinds of representative features, linear prediction cepstrum coefficient (LPCC) and mel-frequency cepstrum coefficient (MFCC), are analyzed and researched on experimentation of Vector Quantization in text-independent speaker recognition. Then, the performances of LPCC and MFCC are compared respectively on computer platform. After analyzing the result of the experiment and the feature for text-independent recognition, we make a solution to the defects of the system. Then long-time spectrum parameter and LPCC or MFCC are blended to be studied in detail to make the feature parameter. Using these technologies, a high recognition rate was made.