Research

Speech enhancement is one of the challenging tasks in signal processing, especially in the case of non-stationary speech-like noise. In this paper we propose a new supervised speech enhancement system that uses Fischer Discriminative Dictionary Learning (FDDL) algorithm to model both speech and noise amplitude spectrum, where the cost function accounts for both "source confusion" and "source distortion" errors. In the enhancement stage, we use sparse coding on the learnt dictionary to find an estimate for both the clean speech and noise amplitude spectrum. In the final stage, a Wiener filter is used to refine the clean speech estimate. Experiments on NOIZEUS dataset using two objective speech enhancement measures: frequency-weighted segmental SNR and Perceptual Evaluation of Speech Quality (PESQ) demonstrate that FDDL outperforms other tested dictionary learning algorithms in the presence of considerable noise (0 dB) for all studied noise types, and in the presence of structured non-stationary noise (ex. car and train noise) for all noise levels.