A bilingual speech interface for New Zealand English to Māori

Abstract

The 'Bilingual Speech Interface for New Zealand English to Māori ' is part of the 'Intelligent Human Computer Interfaces' project under 'Objective 3' of the 'Connectionist-Based Information Systems' programme (FoRST UOO606) Department of Information Science, University of Otago. The project experiments with artificial intelligent knowledge-based engineering methodologies and techniques for designing tools that utilise a hybrid system approach that is both adaptable and flexible to different speakers and languages-namely New Zealand English and Māori. Artificial neural networks, fuzzy rule-based inferencing techniques, genetic algorithms and multimedia based applications form the multiple paradigm approaches for solving real applied problems for speech generation. All are fundamental in the developmental direction of an intelligent human computer interface. Multimedia interfaces to databases use conventional software engineering techniques for the management, access and retrieval of information, therefore, implementing an interface application to access a speech and language database means that a change in computer interaction from manual to automatic control is facilitated between the user and the database. The 'Hybrid Neuro-Fuzzy Speech Recognition System' called HySpeech is based on isolated word recognition for New Zealand English that utilises an English-Māori lexical database to provide an automatic language translator or 'Talking Dictionary'. The current development with version two of HySpeech incorporates the advanced 'Fuzzy Neural Network' models with new 'Learning with Forgetting' algorithms for better speaker adaptive capabilities, an aggregation function provides cleaner phoneme compression and 'Self-Organising Maps' for phoneme and word lists-all to facilitate new language modelling techniques. The language information is housed in a separate database management system designed for HySpeech, it contains; the speakers' characteristics, English and Māori words, phonetic transcriptions and pronunciations, segmental labels, phoneme activation and co-ordinate vectors, and a growing set of digitised Māori speech examples. The interface component of HySpeech thus comprises and text examples from the aforementioned database, interactive dialogues, and a graphical user interface generated environment. Also, experiments in artificial language generation for speech synthesis provide the system with some knowledge and information about the languages. The system could therefore be capable of recreating a form of machine generated speech. Many approaches are available to facilitate synthesised speech, so by utilising the present methodologies and techniques, there is a future solution that can run parallel to the existing direction of HySpeech, for a complete bilingual speech interface. Given that there is presently no operational system that can synthesise New Zealand English and Māori speech, this thesis will also provide some basis to facilitate the further research and development in that area.