Phonetic network representation of phoneme sequences based on CVC-unit concatenation

This research aims to acquire general rules for describing acoustic-phonetic variations of speech in symbolic domain. Concrete methods are presented for acoustic-phonetic variation representation and variation pattern acquisition from a real speech database. The variation patterns are formulated in the framework of a phonetic network by resolving the patterns into CVC-type units. Reversely the network of a speech is composed by concatenating CVC-type units. The acquisition of the network variations is made by estimating the arc's weights of each CVC-type sub-network from the database. This database consists of a phonetically balanced word set containing 1542 words, and all samples of these were labled into compact phonetic segments, called APSEG labels. The resulting arc's weights are filed into a Network-Pattern (NP) table. The sorting and clustering of the sub-network patterns are carried out by revising this table by quantizing the arc's weights.