Speech enhancement using spectral subtraction with wavelet transform

For speech enhancement based on spectral estimation/analysis, an analytic technique by which speech signals can be easily distinguished from noise is desired. The wavelet transform (WT) is an analysis tool for which various types of basis functions can be used. By selecting a proper fundamental wavelet, speech energy can be effectively localized in the space transformed by the WT. In this article, we apply the WT to the spectral subtraction technique, originally defined as using the short-time Fourier transform (STFT). and evaluate the effectiveness of its outcome. Considering the structure of the human voice, we use Gabor and Daubechies wavelets as well as a decaying sinusoid as the fundamental wavelet. The results of computer simulations show that the S/N ratio was improved by the proposed method employing the decaying sinusoid as compared with conventional spectral subtraction. In articulation tests with Japanese nonsense monosyllables, however, no significant difference could be observed.