Speech enhancement with noise parameter estimated by a sequential Monte Carlo method

We present a speech enhancement scheme that is based on sequential time-varying noise parameter estimation and time-varying linear filter. The time-varying noise parameter is estimated within a Bayesian framework by a sequential Monte Carlo method. The method approximates posterior probabilities of speech and noise parameters by a set of samples and estimates the time-varying noise parameters by minimum mean square error estimation over these samples. The time-varying filter can make use of the masking properties of human auditory systems. The proposed speech enhancement scheme can work in non-stationary noise. Experiments were conducted in various non-stationary noise situations, and results showed that the method could have improved performances as compared to some alternative methods.