Hierarchical representation of speech in symbolic domain and fine phonetic labeling

This paper discusses and summarizes the phonetic representation of speech in symbolic domain. A hierarchy is introduced to make clear the levels of descriptions for the process from the orthographical description to the acoustic domain. A representation method by network with APSEGs (acoustic-phonetic segments), which we have already proposed, corresponds to the level close to the acoustic domain, and therefore is useful for resolving nonlinearity of phoneme sequences mapped onto the acoustic domain. The level of the network representation is divided depending on the acoustic compactness of APSEGs. In the latter part of the paper, a concrete procedure is presented to construct a more fine-labeled speech database, based on the former discussion. Some sorted results of the labeling are shown for a phonetically balanced word set.