Title: Recurrent Neural Network and Bionic Wavelet Transform for speech enhancement

Authors: Talbi Mourad, Salhi Lotfi, Abid Sabeur, Cherif Adnane

Addresses: Faculty of Sciences of Tunis, Laboratory of Signal Processing, University Campus, 2092 El Manar II, Tunis, Tunisia. ' Faculty of Sciences of Tunis, Laboratory of Signal Processing, University Campus, 2092 El Manar II, Tunis, Tunisia. ' Faculty of Sciences of Tunis, Laboratory of Signal Processing, University Campus, 2092 El Manar II, Tunis, Tunisia. ' Faculty of Sciences of Tunis, Laboratory of Signal Processing, University Campus, 2092 El Manar II, Tunis, Tunisia

Abstract: This paper deals with speech enhancement using Bionic Wavelet Transform (BWT) and Recurrent Neural Network (RNN). Indeed, it describes a new technique, which removes additive background noise from noisy speech. This technique can be divided into two stages: the application of BWT to the speech signals and the application of an Elman neural network to find an optimal thresholding set to remove related noise wavelet coefficients. Simulation results obtained from computation of the Signal to Noise Ratio (SNR) and the Mean Opinion Scores (MOSs) show good performance of the proposed technique in comparison with many other methods.

Keywords: scales; envelope; perceptual models; cochlear filter bank; bionic wavelet transforms; mean opinion score; recurrent neural networks; speech enhancement; background noise; noisy speech; simulation; signal-to-noise ratio; SNR.

DOI: 10.1504/IJSISE.2010.035002

International Journal of Signal and Imaging Systems Engineering, 2010 Vol.3 No.2, pp.136 - 144

Accepted: 18 Jun 2010
Published online: 31 Aug 2010 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article