Speech analysis method based on source-filter model using multivariate empirical mode decomposition in log-spectrum domain

In this paper, we propose a novel method for speech analysis using multivariate empirical mode decomposition based on the source-filter model. The proposed method decomposes log-spectrum of a speech signal into two groups of the summed intrinsic mode functions (IMFs): the summation of first group contains fine structure corresponding to log-spectrum of the glottal source and the summation of the second group contains spectral envelope corresponding to frequency response of the vocal-tract filter. In the proposed method, the two groups can be automatically determined by using IMF classification based on autocorrelation of multivariate IMFs. We evaluate the proposed method in comparison with linear prediction (LP)-based and cepstrum methods to confirm its abilities. The results reveal that our proposed method can automatically and correctly separate two groups of IMFs, these of source and filter, in log-spectrum representation and effectively analyze speech signal as the same as abilities of LP-based and cepstrum methods while they must be regulated depending upon conditions of usage.