Two-input Two-output Speech Enhancement Based on Speech　Absence Probability Calculated from Two Input Signals

摘要

Retaining good speech intelligibility in sound communication systems in the presence ofnoise is an important issue. Signal processing methods to suppress unwanted noises andenhance the target signal, which are collectively designed as Speech Enhancement (SE),have therefore been studied actively. To facilitate speech comprehension in noisyconditions, two-channel SE, which can preserve binaural perceptual cues for soundlocalization in the two outputs, is much preferred because the perceptual gain by humanbinaural processing up to around 8 dB can be expected along with improvement of theSNR by digital signal processing itself. This paper presents a new two-input two-outputbinaural speech enhancement (BASE) method, which is designed upon the concept ofspeech presence uncertainty characterized by speech absence probability (SAP). In fact,this probability quantifies the degree to which the desired signal is absent. The SAP iscalculated using the a priori SNR and a posteriori SNR with integrated two-channel inputsignals. The target signals at two channels are finally enhanced by this estimated SAP.Comprehensive experiments in various noisy conditions are conducted to demonstrate theeffectiveness of this proposed binaural speech enhancement method.