Linear prediction analysis of speech signals in the presence of white Gaussian noise with unknown variance

Abstract

A simple method is presented to compensate for noise effects before performing linear prediction analysis of speech signals in the presence of white noise with unknown variance. The method determines a suitable bias that should be subtracted from the zero-lag autocorrelation function, rather than deriving the exact noise variance. The resulting linear prediction filter is guaranteed to be stable, since the bias used is always smaller than the minimum eigenvalue of the autocorrelation matrix. In addition to a comparison with other methods, the proposed method is examined from various viewpoints, including the degree of formant intensity, signal-to-noise ratio (SNR), deviation of compensated spectra and objective distortion measures. The improvements observed across a data set, consisting of four sentences uttered by six speakers, indicate that the compensated spectra measured with low SNRs are comparable to the uncompensated counterparts measured with approximately 5 dB higher SNRs.