Speech classification for enhancing single channel blind dereverberation

Several single channel dereverberation techniques exists that enhance the harmonic properties of voiced speech, or utilise a signal model of unvoiced speech. This paper demonstrates how existing speech dereverberation methods can be improved by classifying speech into voiced, unvoiced and silent segments. Methods that enhance the harmonic features of voiced speech can benefit from enhancing only voiced segments, compared to using the entire signal. Removing silent frames from the input can additionally benefit dereverberation methods. Additional features that can be used for dereverberation, signal entropy and minimising the energy of silent periods, are introduced and show good performance. However, speech classification is more difficult for reverberant speech than clean speech. The performance of a number of different classification measures are compared in a reverberant environment. It is shown how performance degrades with increasing reverberation, but some classifiers do hold their performance better than others. The accuracy of the estimation of a dereverberation filter parameter using various signal features are compared. In addition, several signal features can be combined into one cost function. This shows promise in giving improved overall estimation accuracy, by looking at and enhancing a richer set of speech features.