Computationally efficient generic adaptive filter (CEGAF)

Abstract

Enhancement to clean speech from noisy speech has always been a challenging issue for the researcher’s community. Various researchers have used different techniques to resolve this problem. These techniques can be classified into the unsupervised and supervised approaches. Amongst the unsupervised approaches, Spectral Subtraction and Wiener Filter are commonly exploited. However, such approaches do not yield significant enhancement in the speech quality as well as intelligibility. As compared to unsupervised, supervised approaches such as Hidden Markov Model produces enhanced speech signals with better quality. However, supervised approaches need prior knowledge about the type of noise which is considered their major drawback. Moreover, for each noise type, separate models need to be trained. In this paper, a novel hybrid approach for the enhancement of speech is presented to overcome the limitations of both supervised and unsupervised approaches. The filter weights adjustment on the basis of Delta Learning Rule makes it a supervised approach. To address the issue of construction of new model for each noise type, the filter adjusts its weights automatically through minimum mean square error. It is unsupervised as there is no need of estimation of noise power spectral density. Various experiments are performed to test the performance of proposed filter with respect to different parameters. Moreover, the performance of the proposed filter is compared with state-of-the-art approaches using objective and subjective measures. The results indicate that CEGAF outperforms the algorithms such as Wiener Filter, supervised NMF and online NMF.