Speaker Recognition Based On Factor Analyzed Probability Statistic Models

The object for speaker recognition is to determine the identity according to some given speech. This is a biological character recognition technique based on speech. It has wide application prospects in security, justice, military affairs, finance and services. The research of speaker recognition has been carried out all over the world because of its promising role in the information society.For different aim, speaker recognition can be classified as speaker identification and speaker verification. This paper is focused on text-independent speaker identification based on the combination of Factor Analysis (FA) and probability statistical models. Speech endpoint detection and speaker models with their training algorithm for handling intra-frame correlation problem are investigated in detail.Firstly, hidden Markov models (HMM) and Gaussian mixture models (GMM) based on probability statistical speaker models are discussed in detail. The definition of these two models is given and their training algorithm which is the most important issue for these recognizer. So, Baum-Welch algorithm based on Maximum Likelihood Estimation (MLE) is introduced.Then, a method of noisy speech endpoints detection based on Approximate Entropy (ApEn) is proposed. To improve performance of endpoints detection in noisy environment is a significant research in automatic speaker recognition (ASR), especially under the kinds of fact noisy environment. The performance of conventional endpoints detection methods based on short-time energy and zero-crossing rate is unsatisfied in the environments with lower signal-to-noise ratio (SNR). Approximate entropy is a new statistical method of complexity measurement for the research of the time series complexity. It changes little with the variety of the data length and has strong anti-interference ability. The simulation results indicate that the method has high performance in endpoints detection even in some low SNR circumstances.