Speech enhancement using voiced/unvoiced classification

This paper describes a new speech denoising system that exploits both time and frequency behavior of voiced and unvoiced segments and applies in context the appropriate speech denoising technique. This system separates the speech signal into three components a voiced component, an unvoiced component and silence component. A pitch synchronous wavelet transform is used to enhance the voiced component by estimating the signal in term of its spectral content around its harmonic frequency which represent the most active regions in the spectrum, and attenuating the background noise in the inter-harmonic bands. For the unvoiced and silence segments, the denoising procedure consists of applying thersholding in some selective bands and uniformly to all wavelet packet coefficients. The performance of the proposed system is evaluated on speech signal with real world noise, added to it, at various SNR. Experimental results show the accuracy of the proposed approach especially in low signal to noise ratio.