Factors influencing intelligibility of whisper in joint time-frequency domain based on real-valued Discrete Gabor transform

Abstract

Intelligibility improvement of the enhancement algorithm proposed in our previous paper is evaluated in this paper. Joint time-frequency spectrum of the noisy whisper and clean whisper are computed based on real-valued Discrete Gabor transformation respectively. The Gaussian window is used as the kernel function which assures the best utilization of the information area for it possess the smallest product of effective duration. Different ranges of the estimated spectrum were evaluated also to test its impact on the intelligibility of enhanced whisper. Experiment results show that good intelligibility improvement is got by the proposed algorithm and the values of DGT coefficients less than 2c(m,n) is important for improving the intelligibility of whisper.