Speech enhancement by noise driven adaptation of perceptual scales and thresholds of continuous wavelet transform coefficients

Abstract

This paper focuses on employing adaptive scales for computation of perceptually scaled continuous wavelet transform coefficients (CWT) and adaptive thresholding of these coefficients for speech enhancement. The adaptive scales and thresholds both were decided on the basis of the noise level of the noisy speech signal. The CWT coefficients were scaled perceptually and the proposed algorithm suggests selection of number of scales required for analysis on the basis of noise level. The CWT coefficients were then thresholded and for this a novel method of generating adaptive thresholds that too depends on the noise level of the noisy signal has also been proposed. Speech signals were acquired from the TIMIT database and evaluation of the proposed method is done by corrupting these signals by white Gaussian noise (at −10, −5, 0, 5, 10, 15 and 20dB SNRs) and four real world noises (each at 0dB SNR); pink, babble, car interior and F16 cockpit noise from the NOISEX-92 database. Enhancement results are compared on the basis of signal to noise ratio (SNR), segmental SNR (SSNR), spectral distortion (SD) and perceptual evaluation of speech quality (PESQ).Results of the proposed method are evaluated against Ephraim Malah filtering, Stein’s unbiased risk estimate (SURE) thresholding of bionic wavelet transform (BWT) coefficients (BWT-SURE), Wiener filtering (WF), perceptually scaled wavelet packet transform (PWT), multi-model WF and multi-model sparse code shrinkage (MultiSCS) enhancement methods. For the white Gaussian noise case, at all noise levels, SNR and SSNR of the proposed method were better than all the methods under comparison. SD and PESQ results were lower than multiSCS method at 10dB SNR but better at 15dB and 20dB SNRs. For the babble noise case, the obtained results were lower than Ephraim Malah but better than BWT-SURE. SNR and SSNR results for the cockpit noise were comparable with Ephraim Malah and BWT-SURE while for the pink noise case, the proposed method gives the best results.