Speech enhancement using a geometric decision-directed approach

The a priori signal-noise-ratio (SNR) estimator plays a very crucial role in the performance of a noise reduction system. The decision-directed (DD) approach, which is the most widely used technique for estimating the a priori SNR, suffers from one-frame delay bias when following the a posteriori SNR. Many modifications of the DD approach in the literature focus on accelerating the tracking speed by taking advantage of the amplitude spectra but making the phase information neglected. In this paper, a geometric DD approach is proposed to find a better estimate of the a priori SNR by making use of the magnitude spectra as well as the phase information. After representing the signal spectra geometrically in a complex plane, the proposed approach estimates the a priori SNR by incorporating the phase difference between the noise and clean speech spectra into the SNR estimate of the current frame without requiring the orthogonality assumption. The Analysis of the tracking performance indicates that the proposed method can basically avoid the one-frame delay bias of the traditional DD approach, and the evaluation results show that the proposed algorithm performs better than the traditional DD, TSNR and the recent MDD approaches in different the noise conditions and input SNR levels.