Verbal Information Verification System

As an important branch in the area of pattern recognition, speaker verification has attracted attention from more and more scientists since the last twenty years of the last century. There are two way to perform speaker verification: text-dependent and text-independent. Currently, Gaussian Mixture Model-Universal Background Model based speaker verification, dominates the field of text-independent speaker verification. Unfortunately, due to the regardless of content and language information, this kind of system has its limitation when applied to commercial tasks. Text-dependent speaker verification verifies a speaker with the content of utterances. This method will get a better performance than text-independent and can be widely used in commercial tasks. So the research work in this thesis is focus on this method.Text-dependent has a close relation to speech recognition, so we first introduce how an HMM model with well performance is trained. Vocal tract length normalization is used in this thesis and the HMM trained is used in text-dependent speaker verification.This thesis gives introduction on speaker verification and verbal information verification. For speaker verification, we introduced HMM, Vector Quantity and Dynamic Time Warping algorithms. Scores from these algorithms are merged and we finally get a good performance. Discussion on Verbal Information Verification (VIV) is the key point of the thesis. Utterance verification is a general solution for this technology. In this paper, confidence measures based on multiple levels of acoustic layers and models are proposed to give the verification scores. And based on the HMM model training experiments, we introduced a VIV system based on weighted subword score.At the end of the thesis, we combined the SV and VIV to make up a whole system of text-dependent speaker verification system. Its feasibility is proofed by experiment results.