Radial Filter Design Using Kaiser Window Polynomials for Speech Source Separation

Abstract

Separation of coincident speech sources using microphone arrays is a challenging problem. Radial filters can be designed in the spherical harmonic domain to separate coincident speech sources effectively. Limited studies have been conducted on polynomial based near-field radial filter design in the spherical harmonic domain earlier. In this paper, a radial filter design method using a frequency domain Kaiser window function is proposed. The magnitude response of the Kaiser window function exhibits sharp notches, making it a good candidate for the design of radial notch filters. The polynomial expansion of the Kaiser window can be utilized for modeling the spherical array beamformer output. Improved white noise gain and directivity index are noted in the radial filters designed using the Kaiser window function. The radial filter is also used to separate co-incident speech sources. Experimental results on coincident speech source separation indicate better LSD, PESQ, and SDR scores than existing approaches.