Kalman filter with linear predictor and harmonic noise models for noisy speech enhancement

This paper presents a method for noisy speech enhancement based on integration of a formant-tracking linear prediction (FTLP) model of spectral envelope and a harmonic noise model (HNM) of the excitation of speech. The time-varying trajectories of the parameters of the LP and HNM models are tracked with Viterbi classifiers and smoothed with Kalman filters. A frequency domain pitch estimation is proposed, that searches for the peak SNRs at the harmonics. The LP-HNM model is used to deconstruct noisy speech, de-noise its LP and HNM models and then reconstitute cleaned speech. Experimental evaluations show the performance gains resulting from the formant tracking, harmonic extraction and noise reduction stages.