Application of multiple orthogonal window spectrum estimation in speaker recognition

Abstract

For speaker recognition systems, short-time spectrum of speech signal is obtained by using windowed discrete Fourier transform (DFT) in feature extraction. Although windowed DFT can reduces spectral leakage, variance of the spectrum estimation remains high, which reduces the stability of spectrum estimation. Multiple orthogonal window spectrum estimation(referred Multitapering) method, which can not only reduces spectral leakage but also reduces the variance of the spectrum estimation, has more stable performance of spectrum estimate, is utilized in this paper. After how number of windows affects performance of spectrum estimation is studied, the performance of speaker recognition system is also tested in noisy environment. The results show that multiple orthogonal spectrum estimation method has more stable performance and better noise robustness than Hamming windowed DFT.