Hybrid Method for Speech Enhancement Using α-Divergence

A hybrid method for speech enhancement based on Non-Negative Matrix Factorization (NMF) and statistical modeling is presented for using speech and noise bases with online updating is proposed. In the presence of nonstationary noises, template-based approaches have shown better performance when compared to statistical modeling but these approaches depend on a priori information. To overcome the drawbacks of these approaches, a hybrid method is developed. The performance of the proposed method is further improved by considering speech bases as well as noise bases. In terms of Source-to-Distortion ratio (SDR) and Perceptual Evaluation of Speech Quality (PESQ) the proposed method have outperformed the traditional algorithms in nonstationary noise environment conditions.