A modified spectral scaling technique is proposed for enhancing speech corrupted by additive noise. Spectral scaling is a frequency domain noise reduction technique where FFT frequency samples are attenuated to a degree dependent on their signal to noise ratio, so that samples with a low signal to noise ratio are heavily attenuated. This technique reduces the level of noise, but the residual consists of a 'musical noise', which can be annoying to listen to. The noise reduction is improved by the technique described in this paper. For each frame, an estimate of the linear prediction (LP) spectrum is found, and is used to determine a spectral weighting function. The spectral weighting function is incorporated into a spectral scaling algorithm to further attenuate noise in spectral regions where the effect on speech quality is not perceived. The LP spectrum of the speech is estimated by applying LP analysis to noisy speech which has first been subjected to spectral scaling. Peaks that are caused by noise are removed from the resulting LP spectrum. Informal listening tests have shown that this method can significantly reduce the level of musical noise without affecting the speech quality, for input signal to noise ratios of 6 dB and higher.
Bibliographic reference. Crozier, P. M. / Cheetham, B. M. G. / Holt, C. / Munday, E. (1993): "The use of linear prediction and spectral scaling for improving speech enhancement", In EUROSPEECH'93, 231-234.