Modeling vowel normalization and sound perception as sequential processes

Abstract

This study constitutes the first attempt at combining vowel normalization procedures with the linguistic perception framework of stochastic Optimality Theory and the Gradual Learning Algorithm (Boersma, 1998). Towards this end, virtual learners possessing different normalization procedures, and a control learner with no normalization procedure, were trained to perceive Brazilian Portuguese and American English vowels. The only parameters fed into the model were the first two formants of the vowels. The simulations assume that: (i) vowel normalization occurs prior to vowel categorization, (ii) each learner has acquired a different normalization algorithm at the time of the simulations, and (iii) perceptual learning takes place when mismatches between the word intended by the speaker and the word perceived by the virtual listener occur (Escudero & Boersma 2004). Our results show that learners equipped with normalization algorithms outperformed the control learner, obtaining accuracy scores from 20 to 30% higher. When equipped with vowel-extrinsic procedures, learners managed to reach more than 90% of correct responses. Thus, a formal model in which normalization and sound perception are implemented as two sequential processes delivers the expected results, as it accurately models vowel perception even when the training and testing sets have speakers with different ages and gender.