A common approach to improving the performance of a HMM based recogniser in noise is to apply Spectral Subtraction(SS) whereby an estimate of the noise spectrum is subtracted from the input signal spectrum. By using SS we obtain a signal with better features and lower variability. However, over-estimation and flooring make Spectral Subtraction a non-linear compensator and hence the noise level in the compensated system is reduced at the expense of introducing distortion into the speech signal. In this paper, we describe a noise compensation scheme in which SS is integrated into the Parallel Model Combination framework. This scheme allows a set of HMMs trained on clean speech to be compensated for both the effects of the noise and the non-linearity caused by the spectral subtraction. The paper presents an evaluation of this integrated SS-PMC approach using the Noisex 92 database. The results show that this mixed scheme outperforms the recognition performance of conventional Spectral Subtraction. For example, for the Lynx Helicopter noise at Odb SNR, standard fixed variance HMMs give 32% correct, adding SS gives 88% correct and the combined SS-PMC scheme gives 100% correct.

Keywords: Speech recognition, noise, adaptation, spectral subtraction.