A brief review on advancements in Kalman filtering and phase based modulation domain speech enhancement

Abstract

In the present paper, Kalman Filter based speech enhancement algorithms have been studied. Starting from the basic Kalman Filter approach to enhance signal to noise ratio against conventional wiener filtering, to the recent modulation domain Kalman filtering that is based on phase of the speech signal for tracking the phase of speech along with logarithmic spectra of noise as well as speech, the improvisation has been carefully observed and presented here. In the modified Kalman filtering algorithm, for the reconstruction of speech signal, speech phase posterior is utilized for developing an improved phase spectrum of the speech. Kalman filter is operated in two steps, one is to model temporally correlating inter frames of speech and logarithmic spectral-amplitudes of noise, where as the second models their nonlinear relations, assuming speech and noise will get added in complex STFT domain. This method is assessed using speech intelligibility and quality metrics, over a range of SNR values with various types of noise. The performance measures highlighted the consistent enhancement in quality of speech over conventional algorithms used for speech enhancement.