Instantaneous voiced/non-voiced detection in speech signals based on variational mode decomposition

In this paper, a variational mode decomposition (VMD) based method has been proposed for the instantaneous detection of voiced/non-voiced (V/NV) regions in the speech signals. In the proposed method, the VMD is applied in iterative way with specific input parameters. Firstly, the VMD decomposes the speech signal into two components, then, the VMD is applied successively on one of these two components based on suitably defined convergence criteria. It has been shown that the VMD applied in iterative way behaves as a low-pass filter and after convergence it provides separation of the fundamental frequency (F0) component from the speech signal. The envelope of the F0 component of the speech signal has been obtained using an analytical model based on single degree of freedom (SDOF). Automatic threshold has been computed from the obtained envelope in order to detect the V/NV regions in speech signals. The proposed method has been studied on speech signals and the corresponding electroglottograph (EGG) signals from the CMU-Arctic database in different noise conditions obtained from the NOISEX-92 database. Experimental results at various signal to noise ratios (SNRs) are included in order to show the effectiveness of the proposed method compared to the other existing methods for V/NV detection in speech signals.