Single-Channel Speech Dereverberation

We often find audio recordings in which there is too much reverberation but still the recording is important enough to require improvement and post processing to reduce the reverberation. There are already a few methods for so called dereverberation of audio recordings especially for speech signals with particularly varying dereverberation quality depending on different aspects of reverberation. All of them are based on the same fundamental blind deconvolution problem in which neither the original signal nor the impulse response of the system is known. This work discusses a method that combines different approaches based on homomorphic deconvolution and spectral subtraction defined by statistical and power spectral analysis. Optimal settings and algorithmic combinations are found by technical and perceptual evaluation of sound quality and reverberation level.