Quantifying the effect of corpus size on the quality of automatic diacritization of Yorùbá texts

Yorùbá being a tone language requires tone information for the correct pronunciation of words in Text-to-Speech synthesis. Based on standard Yorùbá orthography, such information is held in tone marks, which applied to vowels and syllabic nasals as diacritical markings. However, the tone marks are not always correctly applied in many Yorùbá documents because appropriate input devices for the accurate application of the diacritic marks are not always available. Hence, the absence of tone marks in most written Yorùbá texts presents a major challenge in speech synthesis as the information required for applying the right tone sequences to synthesized Yorùbá speech may not always be available. This study proposes the use of Machine Learning techniques as a basis for the automatic application of tone marks as part of the pre-processing in high level synthesis. Being a resource-scarce language however, there is a lack of sufficiently large Yorùbá corpora for the training of an automatic diacritizer. The study therefore investigated the relationship between corpus size and the quality of automatic diacritization towards estimating the size of corpus required for an ideal level of accuracy.