Speech De-noising using Wavelet based Methods with Focus on Classification of Speech into Voiced, Unvoiced and Silence Regions

This paper presents an improved speech enhancement technique based on wavelet transform along with excitation-based classification of speech to eliminate noise from speech signals. The method initially classifies the speech into voiced, unvoiced and silence regions on the basis of a novel energy-based threshold and then wavelet transform is applied. To remove the noise, thresholding is applied to the detail coefficients by taking into consideration different characteristics of speech in the three different regions. For this, soft thresholding is used for the voiced regions, hard thresholding for the unvoiced regions and the wavelet coefficients of silence regions are made zero. Speech signals obtained from SPEAR database and corrupted with white noise are being used for evaluation of the proposed method. Experimental results show, in terms of SNR and PESQ score, de-noising of speech is achieved using the proposed method. With regards to SNR, the best improvement is 9.4 dB when compared to the SNR of the original (noisy) speech and 1.2 dB as compared to the improvement obtained using one of the recently reported methods.