Speaker
Description
Vowel recognition relies on specific physical properties of sound, known as ``acoustic cues'', that listeners use to identify phonemes. The first two formant frequency values (F1 and F2) have been repeatedly shown to be the primary acoustic cues for vowel categorisation. However, identifying the exact distribution underlying their mental representation remains a difficult task. Current methods, such as reverse correlation, are very time intensive as they typically require thousands of trials. In the present study, we tested whether the method Markov Chain Monte Carlo with people (MCMCp) with adaptive Metropolis-Hastings sampling could provide a more efficient alternative to probe the mental representation of vowels. In a perceptual two-interval, two-alternative forced choice task with synthetic vowels, we showed that vowel-specific regions in F1-F2 space were more precisely estimated using the MCMCp approach than with the reverse correlation method, given the same number of trials. These regions were also congruent with those reported in the production literature. Furthermore, these results do not critically depend on the number of features considered. The use of the MCMCp may therefore prove highly useful for experimental phonetics.