Separation of Two Sound Sources in the Same Direction by Image Signal Processing
Kenji Ozawa, Yuta Koshimizu, Masanori Morise, Shuichi Sakamoto · 2018
This paper proposes a novel method to separate a target sound from another noise in the same direction using a microphone array. Four microphones are arranged to focus the target sound, and their outputs are converted to a spatio-temporal sound pressure distribution (STSPD) image. The focused source forms vertical stripes in the image; thus, their two-dimensional fast Fourier transform (2D-FFT) spectral components are localized as the direct current (DC) components along the spatial axis. Meanwhile, a noise signal forms parabola stripes in the STSPD image and their 2D-FFT components are spread into the target spectrum. However, the leaked DC components can be estimated from the other components, and subtracted from the target spectrum. A computational experiment shows that the noise is suppressed by approximately 25 dB, when the positions of the target and noise sources are known.