Realizing being imitated: Vowel mapping with clearer articulation

摘要

The previous approach to vowel imitation learning between a caregiver and an infant (robot) has assumed that the robot can segment the caregiverpsilas utterance into its phoneme category, where the caregiver always imitates the robot utterance. However, in real situations, the caregiver does not always imitate the robot utterance, nor the robot does have the phoneme category (no segmentation capability). This paper presents a method to solve these issues, a weakly-supervised learning along with auto-regulation, that is active selection of action and data with underdeveloped classifier. To cope with not-always imitation problem, a weakly-supervised learning method is applied that is capable to handle incompletely segmented samples (not perfectly imitated voices). Further, the regulation classifier of the imitated voices is recursively applied in order to select good vocal primitives and to segment caregiverpsilas imitated voices that improve the performance of the classifier itself. The simulation results are shown and the future issues are given.