Spectral multi-normalisation for robust speech recognition

Abstract

This paper presents an improved version of a spectral normalisation based method for extraction of speech robust features in additive noise. The baseline normalisation method was developed by taking into consideration that, while the speech regions with less energy need more robustness, since in these regions the noise is more dominant, the “peaked” spectral regions which are the most reliable due to the higher speech energy must also be preserved as much as possible by the feature extraction process. The additive noise effect tends to flatten the “peaked” spectral zones while the spectral zones of less energy are usually raised. The algorithm proposed in this paper showed to alleviate the noise effect by emphasising the voiced nature of the speech signal by raising the spectral “peaks”, which are “flatten” by the noise effect. The clean speech database is assumed as lightly contaminated, the additive noise is estimated in a frame by frame basis and then used to restore both the “peaked” and the flat spectral zones of the speech spectrum.