Human Speech Model based on Information Separation

Authors: Nobuaki Minematsu

Abstract:

This paper points out that no existing technically-implemented speech model is adequate enough to describe one of the most fundamental and unique capacities of human speech processing. Language acquisition of infants is based on vocal imitation [1] but they don’t impersonate their parents and imitate only the linguistic and para-linguistic aspects of the parents’ utterances. The vocal imitation is found only in a few species of animals: birds, dolphins, and whales, but their imitation is acoustic imitation [2]. How to represent exclusively what in the utterances human infants imitate? An adequate speech model for it should be independent of the extra-linguistic features and represents only the linguistic and paralinguistc aspects. We already proposed a new speech representation [3], called speech structure, which is proved mathematically to be invariant with any kind of transformation. Its extremely high independence of speaker differences was shown experimentally [4, 5, 6]. In this paper, by reviewing studies of evolutionary anthropology and those of language disorders, we discuss the theoretical validity of the new model to describe the human-unique capacity of speech processing.