New AI System Generates Natural-Sounding Yoruba Speech

Researchers have developed TTSYoruba, a rule-based concatenative diphone speech synthesizer specifically designed for the Yoruba language. This system is integrated into the YorubaName.com online dictionary, providing audio pronunciations for personal names. It processes tone-marked Yoruba text and generates speech using a meticulously crafted phonological rule system applied to an inventory of 651 diphone units, encompassing various tonal nuances of consonant-vowel combinations.

The system's architecture addresses critical linguistic challenges unique to Yoruba, such as intricate tonal file selection logic, the disambiguation of nasal sounds (oral /n/, nasalized vowels, and syllabic nasals), and the accurate derivation of contextual rising and falling tones from level-tone inputs. This detailed phonological approach is crucial for generating highly intelligible and natural-sounding speech in a tone-rich language like Yoruba.

Beyond the technical speech synthesis, the project also contributes to Yoruba orthography by formally adopting the caron and circumflex as standard single-vowel contour tone markers. These symbols, already recognized in Yoruba phonological transcription, are now integrated into the system's text normalization pipeline and the WriteYoruba keyboard tool, enhancing the consistency and ease of writing tone-marked Yoruba. A listener study involving 50 participants evaluated the system's performance, with positive Mean Opinion Scores indicating its effectiveness.

This development is highly significant for Africa, particularly for communities that speak low-resource languages. By creating a robust text-to-speech system for Yoruba, it not only preserves and promotes the language but also opens doors for various AI applications, including educational tools, accessibility features, and localized digital content, thereby bridging the digital language divide and fostering inclusive technological advancement across the continent.