Rules, but what for? - rule description as efficient and robust abstraction of corpora and optimal fitting to applications -

Two recent studies are introduced in speech recognition and speech synthesis to reconsider what rules should be looked for spoken language science and technology. To abstract the neighboring characteristics expressed by Ngrams, multi-class composite N-grams have been proposed to model POS characteristics and inflectional forms separately. It is shown that statistical clustering can provide more compact and robust description of word neighboring characteristics than conventional N-grams. For speech synthesis, segmental duration modeling has been examined from the viewpoint of perceptual characteristics of duration changes. A series of perceptual experiments have shown the context dependency of sensitivity to duration change. These two examples respectively illustrate how current rules are interpreted to build scientifically acceptable engineering models and remind us of the deeper scientific meaning and limitation of generalization as a rule.