In this dissertation work, three methods are proposed to improve the performance of speaker recognition systems in noisy environments such as car noise and white Gaussian noise. To construct the automatic speaker recognition (ASR) system robust to environmental noise, we consider both features and system modeling methods.
First, we propose to use prosodic features which represent micro prosody of utterances for speaker recognition. In the case of the background noise, prosodic features and speaking style do not change in contrast with spectral features. The spectral features degrade in noisy environments but the prosodic features are robust. We use the micro prosody which is modeled by segmental pitch contour. Therefore, the codebook is constructed from the segmental pitch contours.
Second, the bootstrap and aggregating vector quantization (VQ) model is proposed. In training procedure, new training sets are made from the original training set by bootstrapping. One codebook is formed from each new training set. Each VQ model from the new training set is used for speaker recognition. Finally, the speaker is identified by aggregating the results of all VQ models. We investigate the unstability of VQ model to apply the bootstrap and aggregating method. Although the bagging VQ model improves the recognition rates significantly, it requires larger memory than the conventional VQ model. Therefore, we propose the probability codebook design method for reducing the additional memory by bagging VQ model. This method uses only one universal codebook for all speakers.
Finally, we propose the independent components analysis (ICA) mixture model for ASR. The first step of the algorithm is to extract the basis vectors from each speaker. The second step is to compute the probability for each ICA class given test data. The third step is to decide the speaker who has the largest probability of ICA. To improve the recognition rates, we assign the number of basis vectors used for e...