Research On The Speaker Identification Algorithms

Speaker Identification is the process of recognize speaker automatically by the information of speaker's voice signal. It can be used in security system, database access, credit card confirmation, telephone trading service of bank and many other situations.This paper focused on parameters model speaker identification. It mainly includes selection, extraction of feature parameters and speaker identification algorithms. The selection of features is to find a group of parameters which can be represent the speaker's voice characters. These features should be insusceptible of environments, robust, keeping acceptable performance for different users and can be used under the normal background noise level. In this paper, linear predictive coefficients (LPC), linear predictive coefficients derived cepstrum coefficients (LPCC) and Mel-Frequency Cepstrum Coefficient (MFCC) is discussed. The performances of these features in the speaker identification are compared.Many algorithms can be used in speaker identification. This paper take the k-nearest neighbor algorithm (k-NN) as the classify method. The performance of different features used k-NN as classifier is compared. Moreover, in order to accelerate the algorithm and save memory, in the model training stage of speaker identification, k-means and fuzzy c-means algorithm is used to reduce the model's size. Compared with the performance of data not reduced model, the results show that these data reduced methods are efficiently in speaker identification.Artificial neural network (ANN) can be used as classifier through training. According the function of classify of ANN, this paper try to apply learning vector quantization (LVQ) of ANN in the speaker identification, and obtain satisfying results.