Marginal energy density over the low frequency range as a feature for voiced/non-voiced detection in noisy speech signals

In this paper, we present a pseudo Wigner–Ville distribution (PWVD) based novel method for the voiced/non-voiced (V/NV) detection in noisy speech signals. The energy distribution of the speech signal on the time–frequency plane is obtained by computing the PWVD coefficients of the analytic speech signal over the low frequency range (LFR). The marginal energy density with respect to time (MEDT) over the low frequency range (LFR) derived from the energy distribution of the speech signal on the time–frequency plane is used as a feature to provide the instantaneous V/NV detection. The experimental results on speech signals from the CMU-Arctic database under white, babble and vehicular noise environments taken from the NOISEX-92 database at various signal to noise ratio (SNR) are obtained to assess the performance of the proposed method. A significant performance improvement in the V/NV detection accuracy is obtained by the proposed method over the existing methods for the V/NV detection under the white noise and babble/vehicular noise environments, respectively.