A Language Independent Approach To Acquiring Phonotactic Resources for Speech Recognition

Building and developing linguistic resources for languages is of prime importance with many areas of application. This paper focusses on a fully automatic approach to the aquisition of a syllable phonotactics for a particular language. In this approach the phonotactic constraints for a language are encoded in a finite-state phonotactic automaton the structure of which can be automatically derived from an example set of well-formed syllables that may occur in the language in question. Such automatic acquisition of phonotactics is achieved through the use of a regular grammatical inference algorithm which is entirely data driven ensuring that it can be applied to any language provided syllable labelled data is available. The approach allows for a rapid and low cost development of phonotactic resources for any language under observation. This makes it an attractive approach for developing phonotactic resources for lesser studied languages in the case where syllable labelled data for a language is available but language specific information required for hand constructing a phonotactics may not. Given that syllable labelled data for a language may not always be available a semi-automatic approach to acquiring a syllable phonotactics from phoneme labelled data without syllable boundaries is also discussed.