SEMIAUTOMATIC LABELING FOR CONSTRUCTING A PHONEME DATA BASE SYSTEM.

This paper presents an improved semiautomatic labeling system which has been developed to construct a fine speech data base. An automatic labeling technique used in the system is based on optimal matching of an input speech sample to reference samples. A major point of improvement this system offers is that it is able to treat connected speech samples by means of a two-stage matching process. The phonetic unit for the labeling is an acoustically compact segment called APSEG (acoustic-phonetic segment). The latter part of the paper discusses speech representation using networks. The acoustic-phonetic variations of the utterances are described by the differences in the label sequences, so that they are represented by a network consisting of the nodes of the APSEG labels.