Study And Implementation Of Speaker Tracking System

With the development of information technology, more and more spoken documents, such as news broadcasts and telephone conversation records, are available. How to find out the appropriate information that people need in them is what the spoken document retrieval(SDR) technologies focus on. Speaker tracking is one of SDR technologies, which can find out the speech spoken by the target speaker in a conversation. Speaker tracking technology is widely used in spoken document processing.In this thesis, a speaker tracking system is implemented, which consists of feature extraction, speaker segmentation and speaker verification. Each part is implemented in two different ways, which are compared through experiments. Research work can be concluded as follows:(1) Linear Prediction Cepstrum Coefficients (LPCC) and Mel-Frequency Cepstrum Coefficients (MFCC) are extracted separately as speaker features.(2) Speaker segmentation based on KL2 distance is implemented. Speaker segmentation based on Bayesian Information Criterion (BIC) is also implemented and improved by variable-length window pre-segmentation method presented in this thesis.(3) Speaker verification systems based on Vector Quantization (VQ) and Gaussian Mixture Model-Universal Background Model (GMM-UBM) are implemented separately.(4) Speaker tracking systems are implemented based on speaker segmentation and speaker verification.(5) Experiments show that the speaker tracking system based on MFCC, BIC and GMM-UBM has the best performance in the themes implemented in this thesis. Its recall rate can reach 93.3% with precision rate 87.5%.