Combined Single-Microphone Wiener and MVDR Filtering based on Speech Interframe Correlations and Speech Presence Probability

For single-microphone noise reduction, a minimum variance distortionless response (MVDR) filter has been recently proposed based on speech correlations of consecutive time frames. This filter is able to keep speech distortion low but compared to conventional approaches achieves less noise reduction. Further, when only having access to the noisy speech, more artifacts in the background noise are audible due to estimation errors of the speech interframe correlations, especially in time-frequency regions where speech is not dominant. Therefore, in this paper we propose to apply the MVDR filter where speech is dominant and the singlechannel Wiener filter otherwise, using a weighting based on the speech presence probability. In addition, we modify the decisiondirected approach to estimate the a priori SNR in a more robust way for short analysis frames. Experimental results show that the proposed scheme achieves a better speech quality compared to the MVDR filter and the single-channel Wiener filter.