ABSTRACT
For pathological voices, hoarseness is mainly due to airflow turbulence in the vocal tract and is often referred to as noise. This paper focuses on the enhancement of speech signals that are supposedly degraded by additive white noise. Speech enhancement is performed in the time-domain, by means of a fast and reliable subspace approach. A low-order singular value decomposition (SVD) allows separating the signal and the noise contribution in subsequent data frames of the analysed speech signal. The noise component is thus removed from the signal and the filtered signal is reconstructed along the directions spanned by the eigenvectors associated with the signal subspace eigenvalues only, thus giving enhanced voice quality. This approach was tested on synthetic data, showing higher performance in terms of increased SNR when compared with linear prediction (LP) filtering. It was also successfully applied to real data, from hoarse voices of patients that had undergone partial cordectomisation. The simple structure of the proposed technique allows a real-time implementation, suitable for portable device realisation, as an aid to dysphonic speakers. It could be useful for reducing the effort in speaking, which is closely related to social problems due to awkwardness of voice.
