Phonetic contrast perception: Predicting human performance from acoustical measures

Abstract

THRIFT is a three-interval forced-choice test of the perception of phonetic contrasts that are embedded in a varying phonetic context. Such parameters as formant frequencies, F0, duration, and intensity were measured for each utterance in a specific recording of THRIFT. Results were used to predict performance of an ideal observer basing oddity decisions on a single acoustic dimension, using (a) a d′ approach with mean differences expressed in relation to token-by-token variability, and (b) an ‘‘oddball’’ approach based on scalar distances among the members of stimulus triads. Reliable perception of some contrasts (e.g., vowel height and place, and consonant voicing) could be attributed to a single acoustic parameter. For other contrasts (e.g., consonant place) no single acoustic parameter was found that could predict reliable perception independent of context. For some contrasts, above-chance performance was predicted for unexpected acoustic parameters that may or may not be specific to this recording. Overall, oddball analysis predicted better performance than d′ analysis, suggesting an advantage for a pre- rather than a postclassification decision strategy. The relative difficulty of contrasts, in terms of predicted perception using single acoustic parameters, parallels the measured performance of deaf children on this test. [Work supported by NIH Grant No. 2PO1DC00178.]