Study On Speaker Identification Algorithm

Speaker recognition technique is one of the speech recognition topics and it is the process of automatical identifying the speacker who is speaking according to the tested speech. Speaker recognition technique has a boradern application prospect, such as the admission control of the secure place or user's identity verification in e-business. According to its various applications, speaker recognition method can be divided into speaker recognition and speaker identification. Speaker recognition can also be divided into text-dependent and text-independent methods. This thesis mainly discusses text-independent speaker verification.A speaker verification system is mainly composed of feature selection and pattern matching. The purpose of feature selection is to find the characterisitic parameters representing a speaker properly. The goal of Pattern Matching is to design valid algorithms that can precisely calculate the distance between a person's characteristic template and the tested speech. The main task of the thesis concludes the following parts:1) Realizing a speaker recognition system using VQ and HMM respectively, and comparing the performances of the two methods;2) Comparing the recognition ratio of LPCC with that of MFCC;3) Addressing the problems of MFCC, united feature parameter of pitch and long-time cepstrum, and analyzing their influnces on the system recognition ratio.4) Training the code book using the MKM algorithm, and adopting the improved method to decrease the waste operation.5) Using the bintree code book which results from the fully searching code book to reduce the computation.6) Proposing a segment weight method of the whole SNR which can improves the identification ratio under the lower SNR circumstance. At the same time, improving the new method to refine its implement on FPGA.