Suppression of combined effect of late reverberation and masking noise for speech enhancement using channel selection method

Abstract

This paper presents a novel channel selection method for suppression of combined effect of late reverberation and masking noises. The proposed algorithm assessed hearing aid users' performance using the Institute of Electrical and Electronics Engineers (IEEE) database corrupted with various noises from NOIZOUS database sentences and evaluated the reverberant mask by varying the threshold from −23 dB to 15 dB. The experiment results show that the performance of the channel selection method improves in highly reverberant conditions (T60 = 1.0 s.) with non-stationary noises at 0, 5, 10 and 15 dB Signal-to-Noise Ratio (SNR). A frame with Signal-to-Reverberant Ratio (SRR) larger than a preset threshold is taken for target speech. This algorithm is capable of providing de-reverberated speech estimate by rejecting the frame of reverberation and non-stationary noise source. This method provides better speech intelligibility with increased SNR and minimises Mean Square Error (MSE) in comparison to other techniques. This also reduces the temporal envelope smearing effects introduced by reverberation and minimises self-masking effect.