Texture analysis approach for improving HMM speech recognition in presence of microinterruptions

A simple yet powerful algorithm for enhancing a signal corrupted by microinterruptions is outlined. The algorithm performs a texture analysis of the peak-based spectrogram image for correcting the damage caused by noise and has been used as a pre-processor in hidden Markov model (HMM) speech recognition. Improvements in accuracy as high as 16% have been obtained with the T120 database.