Showing papers by "Riichiro Mizoguchi published in 1990"

A Support System for Constructing Rule Base for Speech Synthesis by Rule. Automatic Extraction of Synthesis Rules

Abstract: In speech synthesis by rule, various acoustic parameters are determined by rules from the phonetic transcription and then the speech is synthesized. In the past, the synthesis rules have been constructed by an expert on speech synthesis, iterating trials and errors. To construct sophisticated rules efficiently, as well as to improve and maintain those rules, it will be useful to use a computer to support the expert's effort. For this purpose, several functions are required. Among those, this paper discusses especially the method which automatically extracts the synthesis rules from a large number of speech data by inductive inference. In the method, the acoustic parameters, which are determined by the already prepared synthesis rules, are compared with those of the actual speech uttered by a human subject and the difference is described in the form of a “comparison example.” Then inductive inference is performed. In the comparison example, the phonetic environment for the speech data also is described. The concept hierarchy for the attributes concerning the environment is prespecified as the “expert knowledge of phonology,” and is utilized in the inductive inference. Especially for consonants, more than one hierarchical structure is considered, among which the most favorable is employed. The synthesis rules actually were extracted for the CV syllable duration in the word, indicating the effectiveness of the proposed method.