Speech Enhancement Using the IBM Estimate in Real Environments

Speech Enhancement Using the IBM Estimate in Real Environments

In this paper, we estimate ideal binary mask for speech enhancement in real environments contain both noise and reverberation. We try to remove noise and a part of reverberated speech signal using the various estimated IBMs to improve speech quantity and intelligibility. To estimate the three defined IBMs in real environments: IBM-DS, IBM-ER, and IBM-R are used the SVM classifier and the popular features: GFCC, PNCC, MFCC, and RASTAPLP. First for each mask the optimal LC is determined based on best results of STOI criterion then, are obtained the results of SNR-improvement and PESQ value. Overall, the simulated IBMs based on MFCC have the best PESQ results and GFCC-based have the best SNR-improvement rather than other features. It should be noted that for noises which include of speech signals the RASTAPLP has better SNRimprovement results.