Efficient spectrum estimation of noise using line spectral pairs for robust speech recognition
A novel method for estimating the power spectral density of acoustic background noise is proposed. The spectral peak frequencies are approximated by the roots of the P polynomial, which constitute half of the line spectral pairs. The probability distributions of the magnitude values at the spectral peaks are modelled by a mixture of two univariate Gaussian functions, where the Gaussian with smaller mean is considered as noise and the other as speech. The validity of the proposed method is exhibited by the experimental results evaluated on a simple speech recognition task.
