New Research Questions Linguistic Relatedness for Large ASR in Low-Resource African Languages

The development of Automatic Speech Recognition (ASR) for the vast number of low-resource African languages faces significant hurdles, primarily due to the immense data collection requirements. A common approach to mitigate this challenge involves leveraging the linguistic similarities between languages to facilitate cross-lingual transfer from a well-resourced auxiliary language to a low-resource target language. This strategy has shown promise in smaller ASR models, but its efficacy in the context of larger, more complex multilingual ASR systems has remained largely unexplored.

This new research extends the investigation of linguistic relatedness to large multilingual ASR models through a rigorous and controlled experimental design. The study involved six distinct factors, incorporated two Africa-centric corpora, and evaluated four different large ASR models. The primary objective was to determine whether linguistic relatedness could reliably predict gains from cross-lingual transfer in this advanced ASR setting, particularly when dealing with minimal target-language data.

Contrary to expectations based on prior work with smaller models, the findings indicate that pre-adaptation on linguistically related auxiliary languages did not yield any practically meaningful transfer improvements. This was consistently observed across all experimental conditions, even when target-language data was minimal. The study concludes that linguistic relatedness alone may not be a dependable predictor of cross-lingual transfer gains in large multilingual ASR, nor does it appear to be an effective standalone strategy for expanding such models to low-resource languages.

For Africa, where linguistic diversity is immense and many languages lack extensive digital resources, these findings carry significant implications. The reliance on linguistic relatedness as a primary method for bootstrapping ASR capabilities for underserved African languages may be insufficient for large-scale, high-performance models. This research underscores the need for exploring alternative or supplementary strategies to overcome data scarcity and accelerate the development of robust ASR technologies that can effectively serve the continent's diverse linguistic communities. It highlights a critical challenge that must be addressed to ensure inclusive AI development in Africa.