Three-dimensional modelling of speech corpora: added value through visualisation

Collections of annotated spoken language have formed an important basis for the development of speech technology. Their existence has promoted speech analysis research as well as enabled robust synthesis and recognition methods to be developed. However, many complex relationships remain unspecified within a corpus due to a lack of meta-data that describes the raw information in sufficient detail as well as the interrelationships between signals, recording conditions, talkers, etc. A deficit of standards and formats, needed to express complex relationships, has also hindered the potential use and value of available corpora. This paper presents a novel three-dimensional model for exploring temporal as well as atemporal information existing in speech corpora. Examined are the potential benefits that are gained through corpus visualisation during the phases of creation, editing, verification, use, and exploration. The paper suggests that by providing a three-dimensional model of speech data, more of the inherent and potential value of a corpus can be utilised.