Comparison between AR and SVD approaches for speech denoising

For pathological voices, hoarseness is mainly due to jitter and air flow turbulence in the vocal tract. In this paper, speech denoising is performed in time-domain, by means of a fast and reliable subspace approach. A low-order singular value decomposition allows separating the signal and the noise subspace of an appropriate data matrix, obtained from short data frames. The filtered signal is reconstructed along the directions spanned by the eigenvectors associated to the signal subspace eigenvalues, thus disregarding the noise contribution. This approach was found to work well for hoarse voices coming from cordectomised patients, giving enhanced voice quality. Its simple structure allows a real-time implementation, suitable for portable device realisation.