Testing an auditory model by resynthesis

This paper describes the use of a resynthesis strategy in testing an auditory model, specifically a version of the DOMIN model. The steps by which the original speech is processed to produce a reduced, auditory (DOMIN) spectrum are described; subsequently, the reduced representation is used to produce resynthesised speech having the same auditory spectrum. In light ofthis equivalence, the extent to which the original and resynthesised speech are perceptually equivalent is argued to be a good test of the model. We show that a spectral representation in which approximately two-thirds of the FFT frequency components are discarded, but the DOMIN representation is unchanged, can produce resynthesised speech of high intelligibility. We conclude that the DOMIN model retains important information pertaining to the identity of both vowels and consonants. Further, we present evidence showing that testing by resynthesis is superior to the alternative techniques for assessing auditory models.