In order to understand how humans learn speech imitation without access to detailed articulatory data of other talkers, simulated speech acquisition experiments between two virtual agents were carried out with the goal of maintaining the interaction between the two as natural as possible. As an outcome, a novel model of infants’ vowel acquisition is presented. In the experimental setup, a virtual infant learns vowels in interaction with a virtual caregiver: it babbles vowels randomly, the caregiver answers every babble with an utterance that contains the vowel uttered by the infant in addition to other vocalic content, and the infant associates its own productions to the caregiver’s responses. The infant and the caregiver have different vocal tract sizes, and hence the acoustic qualities of the same vowel differ between the infant and the caregiver. The infant learns on line to map acoustic qualities of its caregiver’s speech onto its own vowel articulations, allowing for instant imitation of the caregiver’s vowel sounds when recognized. As opposed to previous computational studies of vowel acquisition, the infant does not need initial mappings, initial vowel primitives, or knowledge of the caregiver’s vowel categories.
Index Terms: speech acquisition, vowel learning, imitation
Cite as: Rasilo, H., Räsänen, O., Boer, B.d. (2013) Virtual infant’s online acquisition of vowel categories and their mapping between dissimilar bodies. Proc. Speech Production in Automatic Speech Recognition (SPASR-2013), 50-55
@inproceedings{rasilo13_spasr,
author={Heikki Rasilo and Okko Räsänen and Bart de Boer},
title={{Virtual infant’s online acquisition of vowel categories and their mapping between dissimilar bodies}},
year=2013,
booktitle={Proc. Speech Production in Automatic Speech Recognition (SPASR-2013)},
pages={50--55}
}