Human listeners are able to perceptually compensate for the effects of reverberation on speech recognition, by exploiting information gleaned from prior exposure to the reverberant environment. We present a computer model of perceptual compensation for reverberation implemented within a hidden Markov model speech recogniser, in which different reverberant speech models are selected depending on the acoustic context preceding a test word. During decoding, observation state likelihoods were computed from two reverberant acoustic models in parallel, and weighted according to the amount of reverberation in the first 500 ms of each utterance. The confusions made by the computer model closely corresponded with those made by listeners in a consonant identification task, and showed a perceptual compensation effect.
Index Terms: reverberation, speech perception, computer model
Bibliographic reference. Brown, Guy J. / Beeston, Amy V. / Palomäki, Kalle J. (2012): "Perceptual compensation for the effects of reverberation on consonant identification: a comparison of human and machine performance", In INTERSPEECH-2012, 1716-1719.