A Dynamic Feature Extraction Based on Wavelet Transforms for Speaker Recognition

In this paper, we try to place emphasis on the dynamic feature extraction and detection of speech signal under the forensic identification practice, which meet the demand of text-independent speaker recognition analysis. As a result of voice changeability, accent diversity and influence by all kinds of environmental factors, voice detection become very difficult. The text-independent speech signals are non-stationary and have important parameters included both in time and frequency domain. Wavelet transforms is proposed because it holds multi-resolution analysis ability and a reasonable wavelet cores is selected. Additionally, a new combined-future-extraction algorithm based on Mallat algorithm of biorthogonal wavelet is proposed to extract the new feature of voice and detect the singular signal in noise environment. Theoretic analysis and simulated experiment results show that new algorithm not only achieves new signal with precisely and available by decomposition and reconstruction to the target signal, but also suits to real areas, real-time application and also can be strengthen. The experimental results are better to verify the above analysis.