Research On Improvement Of Speaker Recognition Algorithms Based On Hand-held Device

The thesis presents and implements a novel Speaker Recognition System on PDA. This system offers voice collection, speaker verification, speaker identification and incremental learning.This system provides many Automatic Speaker Recognition methods and arithmetic, including speech collection, pre-processing, feature extraction, speaker modeling, recognition result gaining and incremental learning. Furthermore, the thesis brings forward and fulfills some new methods and arithmetic in the design and implement of the system:1. Execution time optimization in speaker recognition on PDA. To accomplish real-time verification in a mobile embedded system, the execution time of feature extraction and pattern matching must be limited. This paper designs a highly optimized speaker recognition system on a PDA. And the experiment shows the time reduction from original system to optimized system.2. Feature estimation and selection. The traditional feature extraction seems good to recognize speaker by their voice in small case. In the real world, some dimensions are affected by the environment such as the channel and noise. Based on the feature estimation, we calculate the contribution of each dimension. And we use a selection method to improve the verification rate.3. Model searching improvement. The traditional identification method need to calculate all the speaker models in database and takes a long time. Applying the model clustering method, the system avoids calculating every speaker model in the system. In order to avoid redundant computation of Gaussian function, a new kind of GMM-UBM model is designed based on coefficient adjustment.4. Verification strategy. Based on the result of clustering, we design a verification strategy and the experiment shows this method enhances the performance. We also design a fusion strategy for multi-modules verification tasks.5. Model incremental learning method. In this system, the model incremental learning method is applied. It is used to improve the time robustness of the speaker recognition system. The experiments show this method significantly appeases the time mismatch problem.This work is supported by National Natural Science Foundation of P.R.China(60273059), Zhejiang Provincial Natural Science Foundation for Young Scientist of P.R.China (RC01058), Zhejiang Provincial Natural Science Foundation (M603229) and National Doctoral Subject Foundation (20020335025), National Science Fund for Distinguished Young Scholars60525202, Program for New Century Excellent Talents in University NCET-04-0545 andKey Program of Natural Science Foundation of China 60533040.