Text-Independent Speaker Identification Using an Improved Hidden Markov Model

Abstract

In this thesis, we attempted to build speaker identification system. The purpose of this thesis is to improve the performance of speaker identification system based on hidden Markov model classifier. We proposed an approach in which the coefficients of LPCC or MFCC can be increased, the codebook size can also be increased, and HMM classifier can be trained and tested with multiple codebooks. The implementation of the system is divided into three stages: feature extraction, vector quantization, and training and classification. In feature extraction stage, we studied LPCC and MFCC by which the spectral features of speech signal can be estimated. Then we showed how these features can be computed. In the vector quantization stage, we generated a distinct codebook of size 256 clusters for each speaker model. In the stage of training and classification, we studied the algorithms and implementation of HMM. Then we showed how HMM can be trained to estimate the model parameters using Baum-Welch algorithm and how can be tested using forward algorithm. To evaluate the HMM classifier, we extracted a sample from the Switchboard dataset. The sample consists of recordings of 40 speakers of American English. Experimental results show that the average identification rate of 97.5% has been obtained. Also the results show that the identification rate of LPCCs is better than that of MFCCs. We founded that the system can achieve a stable identification rate with a codebook of size >=256. We compared LDA and MLP classifiers with HMM classifier. Results illustrated that the HMM classifier achieves a superior results. These results conclude that the proposed approach can improve the performance of speaker identification system.