Abstract:
Literature review showed that some languages are extinct while others face extinction.
Extensive studies further revealed that some languages face extinction because of
inadequate digital research and scarcity of natural language processing resources. Yoruba is resource-scarce in the area of speech synthesis and recognition. It lacks research resouces like optical character recognition (OCR), syllabificator and annotated corpus among others.
Yoruba is still under-researched in the area of prosody, needing attention. Consequently,
from literature, some technologies that produced good prosody in certain syllable based
languages were identified. In this study, a Tesserract-based OCR application was developed for Yoruba. This was used to derive sentences from some Yoruba printed texts.The sentences were used to develop a read-corpus. Thereafter, a corpus based prosodic Textto- Speech System (TTS) was developed using unit selection approach. Target costs and join costs were estimated using a system of context-free grammar (CFG) trees and context based binary decision tree whereby Acoustic Space Formulations (ASF) was applied. Viterbi’s search was used to select candidate strings with matching ASF from the speech database while incurring minimum total cost. Using phonetically balanced sentences, a speech database of Standard Yoruba (SY) was developed to support the TTS. A syllable map based SY syllabificator was developed and integrated into the system. A mean opinion score (MOS) perceptive evaluation and semantically unpredictable sentences (SUS) test was carried out on the TTS to respectively assess naturalness and intelligibility. The result compared favourably with existing works in SY TTS. This thesis presents historical background, motivations, objectives, designs, implementation of system and subsystem components. There exist the need to elucidate better search algorithms that could enhance the responsiveness of corpus-based systems and the need to make the system webbased through further research.
