Research On Speaker Recognition And Its Robustness In Noise Environment

Speaker recognition is task of identifying or verifying a speaker through the voice of a given group of speakers. Depending on spoken utterances, speaker recognition can be divided into text-dependent and text-independent. The performance of speaker recognition system under the condition of low noise and low distortion has reached high levels of satisfaction, but the system performs poorly under noisy conditions as acoustic models and test speech can be mismatched. Therefore, improving the robustness of speaker recognition in a high noise environment is a key factor for successful application of the system and it is also a research hotspot.Fundamentals of speaker recognition were firstly presented in details in the paper and two speaker recognition systems based on vector quantization and GMM models were developed in this effort. The first system used three kinds of clustering methods to develop codebooks and then compared the performance of codebooks by the quantization error on the training data and error rate. The second system tested the performance of LPC,LPCC, MFCC and MCC, and found that the feature parameters based on the auditory were more authentic and robust. Moreover a mandarin tone recognition system was developed in chapter three when we presented the HMM model.Considering the poor robustness of speaker recognition system in a noise environment, this paper presented a feature extraction algorithm based on wavelet decomposition of the autocorrelation sequence and a speaker recognition based on the confidence analysis. As autocorrelation sequence is robust to stationary noise and slow-varying noise, we applied multi-level wavelet decomposition on the relative autocorrelation sequence in the feature extraction algorithm. And in the speaker recognition based on confidence analysis, we used the confidence to measure the robustness of feature parameters and proposed a new method called CBTM to get the confidence of each MFCC component. The CBTM evaluated the confidence of all MFCC components disposed by the Mel spectral subtraction through a confidence transformation matrix, and found that reducing the impact of component with low confidence on the output probability through weighting the GMM variance can help to improve the robustness. The experiments demonstrated that the confidence analysis can further reduce the baseline system's error rate.