Roel Smits Rene Collier
Inst. for Percept. Res., P.O. Box 513, 5600 MB Eindhoven, The Netherlands
Two male Dutch talkers produced two tokens of each of 20 stop-vowel syllables (/b,d,p,t,k/ followed by /a,i,y,u/). The release bursts were separated from the voiced parts and four types of stimuli were created: Burst-only stimuli (BO), burstless stimuli (BL), stimuli with cross-spliced bursts, where burst and transitions indicate conflicting place-of-articulation information (CS), and original utterances (OR). These stimuli were presented to 20 subjects for identification. Results show the well-known vowel-dependent trade-off between burst cues and transition cues. An attempt was made to predict the confusion matrices from acoustic properties of release bursts and formant transitions using Luce's similarity-choice model. Most of the variance (89%) in the confusion matrix for the BO condition was explained using a spectral tilt measure and a spectral compactness measure. Much of the variance (80%) for the BL condition was explained using frequencies of F2 and F3 at voicing onset and of F2 in the vowel. Using the same burst and transition measures for the CS condition the explained variance dropped to 40%. Preliminary results show better predictions when acoustic measures are used which integrate over burst and transitions.