Sound Source Separation by Spectral Subtraction Based on Instantaneous Estimation of Noise Spectrum

In our previous paper, we proposed a sound source separation method using the two-dimensional fast Fourier transform (2D FFT) of a spatio-temporal sound pressure distribution (STSPD) image that is composed from the outputs of a microphone array. In an STSPD image, vertical stripes are created for a target sound arriving from the perpendicular direction to the array; therefore, its spectral components are concentrated on the spatial direct current (DC) components in the 2D amplitude spectrum. In that study, we estimated the noise DC amplitudes using a deep neural network (DNN), then subtracted them from the observed spectrum to suppress the noise. However, the performance of noise suppression can be improved further. In this study, we estimate the noise DC components theoretically instead of empirically using a DNN. We improved the performance successfully.