Signal enhancement for continuous speech recognition

Abstract

This paper presents a comparison between two parametric methods for Signal Enhancement in order to address the problem of robust Automatic Speech Recognition (ASR). An SVD-based technique (ISE) and a non-linear spectral subtraction method (NSS), have been evaluated by means of the Continuous Speech Recognition system that is used in the ERMIS project. The input signal is corrupted with coloured noise with variable signal-to-noise ratio. It was found that fine-tuning of the various parameters of the enhancement techniques is crucial for efficient optimisation of their performance. Both methods provide significant improvement of the speech recogniser performance in the presence of coloured noise, with the NSS method being slightly better. © Springer-Verlag Berlin Heidelberg 2003.