New MGhana-ST Dataset Boosts Speech Translation for Ghanaian Languages

Researchers have introduced MGhana-ST, a new speech translation dataset specifically designed for four low-resource Ghanaian languages: Ga, Twi (Akuapem and Asante dialects), Ewe, and Fante. This dataset represents a significant effort to address the scarcity of linguistic resources for African languages, which is a major barrier to developing AI technologies that can serve diverse populations across the continent. The initial release includes 16.1 hours of paired speech and English translations, meticulously annotated by 37 native speakers.

The study accompanying the dataset explores the efficacy of multilingual versus monolingual training approaches using a Whisper-small model under conditions of severe data scarcity. The findings indicate that simple flat multilingual training does not universally benefit all low-resource languages, with some, like Ewe and Fante, experiencing performance degradation. This highlights the complex challenges in applying general AI models to highly diverse linguistic contexts and the need for nuanced strategies that consider linguistic distinctiveness and data availability.

A crucial methodological insight from the research is the importance of robust experimental design, specifically the use of multiple seeds for training. An earlier single-run analysis had suggested positive transfer for most languages, but this was found to be an artifact of seed variation rather than genuine transfer. This finding underscores the need for rigorous scientific practices in AI research, particularly when dealing with low-resource languages where subtle differences can have significant impacts on model evaluation and development.

MGhana-ST is being released to the public, aiming to catalyze further research and development in African language speech technology and low-resource speech translation. This initiative is vital for fostering digital inclusion and ensuring that AI advancements are accessible and beneficial to the linguistic diversity of Africa, moving beyond dominant global languages.