An Experimental Study of Distinctive Features Using Speech Recognition Technology

Abstract

The present paper investigates the distinctive features using a recent developed technology for automatic speech recognition. This technology, "Phoneme Environment Clustering" (PEC), is formulated as an estimation of the mapping function from the "phoneme pattern space" which is a vector space corresponding to the phonetic (acoustic) aspect of the segment. The process of the successive splitting of sub-spaces forms a tree structure. Using PEC, we have examined approximately 2, 000 segments from 216 phonemically balanced words uttered by a Japanese male informant. The results show that this tree diagram well corresponds to the phonetic "natural classes". The classification of segments into several groups reflects the differences between distinctive features. The results also show that the features [sonorant] and [consonantal] are separated from the others at the earliest stages of the process of constructing the tree structure. These acoustic hierarchies coincide with feature hierarchies based on articulatory properties in phonological feature geometry.