Reverberant word intelligibility and psychological models of dereverberation.

摘要

Reverberation interferes with the ability to understand speech in small rooms, so most listeners use binaural information to improve intelligibility. This could mean that reverberation is reduced neurologically before the word is identified. For example, each echo contains binaural cues that might be used for localization and echo reduction. These cues are weak and it is difficult to imagine a binaural process capable of reducing thousands of coherent echoes. Overlap-masking explains intelligibility loss due to reverberation in terms of noise masking. It assumes reverberant phonemes endure in time and mask subsequent phonemes. In this case, the binaural system may be able to reduce reverberation based on the uncorrelated signals at each ear. It is also possible that reverberation is not reduced, but binaural information is used by higher level neurological processes to improve intelligibility. These high level processes might gather phonetic information from each ear to improve intelligibility.; The binaural word intelligibility advantage is investigated through a series of word intelligibility tests conducted in reverberant rooms. These measure the intelligibility of phonetically balanced word lists diotically and binaurally to determine the magnitude of the intelligibility difference. In later tests, reverberation is modified to create reverberation-like noise. The reverberation-like noise has similar temporal and spectral properties to reverberation, but does not contain binaural localization cues. This noise is used to test the hypothesis that the binaural word intelligibility advantage is a result of binaural masking release as opposed to binaural localization and echo reduction. The results show that a small binaural intelligibility advantage exists, and that binaural masking release accounts for only a small portion of this advantage. Either localization cues of the echoes are necessary to achieve the advantage or higher level processing is utilizing the binaural information.; A signal processing algorithm is designed to increase reverberant word intelligibility. It is based on the hypothesis of binaural masking release; which did not hold up in the intelligibility tests. Consequently, any advantage gained by reducing overlap-masking is overshadowed by the detrimental effects of the process. The net result is that the processing reduces word intelligibility scores.