Dereverberation Using Multiple-step Linear Prediction Modified with Gammatone Filter Bank

When people use communication equipment in an enclosure room, the speech signal captured by a distant microphone is usually contaminated by reverberation and noise, which will severely degrades the speech intelligibility. After the noisy reverberant speech signal is processed with denoise and dereverberation algorithms, residual “musical noise” has a great effect on the speech intelligibility and auditory comfort level. In this paper, we introduced Gammatone filter banks which were designed to simulate the human basilar membrane response to sound input, and used them to smooth and modify the gain function obtained by spectral subtraction method. Experimental results showed that the new algorithm suppressed the “musical noise” efficiently and consequently improved the speech intelligibility and auditory comfort level.