Using pronunciation data to study word recognition
Mark A. Pitt, Keith Johnson · The Journal of the Acoustical Society of America · 2003
Many of the mysteries of spoken word recognition have evolved out of the observation that pronunciation is highly variable yet perception is amazingly stable. Extreme forms of variation can even result in different phonetic percepts, such as consonant assimilation (e.g., ‘‘green ball:’’ ‘‘greem ball’’) and deletion (e.g., ‘‘and:’’ ‘‘an’’), yet listeners still perceive the intended word. The results from two lines of work will be presented in which we are studying the regularity of these production phenomena to evaluate models of how phonological variants are recognized. In one project we examined variation itself. Phonological and acoustic analyses of phonological variation in the Buckeye corpus of conversational speech were carried out in which we asked questions such as the following: How predictable and consistent is regressive assimilation? and How acoustically similar is the assimilated segment to an intended production of that same segment? In a related project, we examined the listener’s sensitivity to the variation found in conversational speech. Results reveal the complexities of these production phenomena, the challenges models must overcome to account for how variants are recognized, and listeners’ sensitivity to stochastic properties of pronunication variation. [Work supported by NIDCD.]