Discriminating Same or Different speech: human vs machine

We compare the performance of a Bayesian classifier against humans in a same/different speech perception task. The classifier is trained on different sizes of speech segments to separate sounds into perceptually equivalent and nonequivalent categories. We test the classifier with pairs of speech sounds of various sizes and compare its performance with perceptual listening tests on native speakers of Southern British English as well as Greek native speakers who speak English as a second language. We find that the performance of the classifier is comparable to that of human subjects and that the largest-sized speech segments produce the best performance overall. We also find that an edge detector in the 60-800 Hz band and an autocorrelation-based measure of voicing are very important to the performance of the classifier.