Design and implementation of a Luganda speech database for unit selection speech synthesis using phonetic transcription

Abstract

Speech synthesis systems have greatly simplified natural language interaction between humans and computers. A number of such systems exist with several languages for example Indian, English and Arabic, incorporated in them. However, there has been limited implementation in African languages like Luganda. The few that exist produce unnatural and unintelligent speech since they do not take into account the “co-articulation” effects of producing phonemes, phoneme duration or desired pitch variation across the utterance in continuous speech. There is therefore a need to develop a Luganda speech database that uses phonetic transcriptions as a method for unit selection synthesis that can be applied in Text to Speech systems, phonetic analysis, speech analysis and Digital Signal Processing. It reflects on a single speaker speech database with carefully recorded audios under studio conditions consisting of 1578 phonetically balanced Luganda utterances. We predict a high usability for semi-literate and visually impaired Luganda speakers when applied to speech synthesis systems but also for further learning. More improvements can be through focusing on the speaker traits such as gender, age and origin of the voice talent person. The same approach can also be extended to other indigenous African languages other than Luganda to build systems that can help the semi-literate and visually impaired people.