The perception of scale in vowels

The resonating properties of many objects provide acoustical correlates which can be used to gain information about the objects. The acoustic signal provides not only shape information (what the sound means) but also size information (how small/big the object is relative to the population). A signal processing algorithm able to isolate both shape and size information is the Mellin transform. It is posited that such a transform is applied to all sounds at an early point in the auditory system (Irino and Patterson, 2002). Our ability to tell what vowel was spoken (vowel normalisation), despite gross waveform changes due to different vocal tract lengths and larynx size differences across sex and age, shows we are able to extract the invariant vowel qualities (shape information). We are also able to tell whether the speaker was a man, woman or child (size information). We manipulated vowel sounds to have the acoustical properties of different size speakers scaled way beyond the usual range of variation in the population. Listeners were able to both identify the vowels and extract information about the size of the speaker of these scaled vowels. Our results are unlikely to be due to some speech-specific learnt statistical correlation between pitch and formant frequency as the learnt association should fail outside the training set region. The huge range over which listeners were able to extract shape and size information suggests rather the operation of some active pre-processing transform that is applied to all input sounds.