Automatic extraction of fundamental frequency control rules by statistical analysis

This paper aims at the improvement of the naturalness of Japanese synthetic speech and proposes a method of extracting automatically the rules for controlling the voice fundamental frequency (written as F₀). The proposed method is composed of two steps. (1) The F₀ time-series pattern of a sufficient amount of speech data is represented by parameters under Fujisaki's model. (2) The F₀ control rule that estimates die parameter values from the language information is extracted by statistical analysis. The proposed method is applied to 200 Japanese sentences read by a speaker, and the relation between the language information and the parameter values is derived by analyzing the obtained F₀ control rules. The following properties are identified. (1) The phrase command diminishes when the number of morae in the preceding phrase decreases. (2) The accent component is reduced when the number of morae in the higher-pitched part of the accented phrase is larger. Relation (2) is a refinement of knowledge already obtained by the analysis of a small number of samples. Thus, it is shown that adequate F₀ control rules can be extracted automatically by the proposed method.