On the link between acoustic cue distributions and categorization dependencies
Roel Smits · The Journal of the Acoustical Society of America · 1999
The HICAT model of hierarchical categorization [R. Smits, J. Acoust. Soc. Am. 103, 2980(A) (1998)] explicitly models dependencies in four-alternative forced-choice categorizations involving two binary distinctions. A philosophy has been developed which predicts which dependencies are likely to occur for a given pair of phonetic distinctions on the basis of the acoustic distributions of relevant cues in natural utterances. It is argued that, because speech perception happens under severe time pressure, listeners try to minimize categorization complexity while maximizing categorization accuracy. For certain phonetic distinctions, the cue distributions allow listeners to use simple, independent categorization strategies while still performing close to optimally. Severe coarticulation, however, necessitates a more complex strategy involving categorization dependencies which reflect dependendencies in the cue distributions. This philosophy was tested experimentally. Spectral locations of fricative and vowel resonances in syllables /si, sy, Si, Sy/ were measured on a large set of naturally spoken tokens. Based on the resulting distributions it was predicted that listeners’ fricative categorization will be dependent on the vowel categorization. Next, a categorization experiment was run using a two-dimensional fricative-vowel continuum. The dependencies inferred from the categorization data using the HICAT model will be compared to the predictions.