Speech enhancement using a variable suppression rule in Hilbert domain

We describe a speech enhancement system which combines a variable, input adaptive noise suppression rule with a recently developed spectral analysis framework in Hilbert domain. The variable suppression rule is an extension to a formula which encompasses well known noise reduction algorithms such as power subtraction and Wiener filtering. Time-varying parameters which are based on the input signal to noise ratio and its spectral shape are embedded in this formula. The framework consists of using Hilbert transform in sub-bands in conjunction with wavelet packet decomposition. This spectral analysis accounts for perceptual features and it was shown to be more effective than the common Fourier transform. The experiments show speech quality improvement in terms of perceptual measures.