MOOSE: management of otago speech environment

With the advent of spoken language computer interface systems, the storage and management of speech corpora is becoming more of an issue in the development of such systems. Until recently, even large corpora were stored as individual text and speech files (for example, [18]), or as a single, monolithic file (for example, [1]). The issues involved in management and retrieval of the data have been, to a large extent, overlooked. Initially, the Otago Speech Corpus files were stored in a nested directory structure, including speech and text. Although each item of data could be retrieved, there was no relationship between them. Therefore, a phone extracted from a spoken word had no direct relationship back to that word. Furthermore, with no management system, a new user of the corpus was required to become familiar with a cryptic naming scheme and the nested directory structure before the corpus was useful. Relational database management systems (RDBMS) are proposed as an ideal tool for the management of speech corpora [5]. Relationships between words and phonemes, and the realisations of these, can be stored and retrieved efficiently. RDBMS may be constructed with various levels, to store speaker, language, label transcription, and phonetic information, plus speech as isolated words, and derived segmented units. In this manner, the data becomes a complete speech corpus. An implementation of such a system is presented to manage the Otago Speech Corpus, currently called Management Of Otago Speech Environment (MOOSE). The ability of the MOOSE to be applied to other corpora is currently under investigation.