A New Approach to Offline, Community-Driven AI for African Languages

This paper introduces "Model as a Library" (MaaL), a novel software architecture designed to bring AI-powered services to low-resource African languages. Traditional large language models often fail these communities due to their reliance on web-scraped data that doesn't capture the vast dialectal and regional variations of African speech. MaaL aims to overcome this by enabling offline, on-device AI capabilities.

MaaL packages small, community-sourced speech models as versioned, on-device dependencies. This allows for structured data collection without the risk of generative hallucinations, which is crucial for populations poorly served by current language models. Instead of relying on vast, standardized text corpora, MaaL's vocabulary is directly enrolled from a small number of example recordings provided by the speakers themselves at the point of deployment.

The core mechanism of MaaL is keyword spotting, which transforms closed-vocabulary text forms into voice forms that can be filled and submitted entirely offline. The authors propose integrating MaaL schemas into existing digital form tools, offering a low-friction pathway to voice-first, offline data collection for low-literacy populations already using these tools. This system design paper outlines the concept, mechanism, and analytical feasibility, while also identifying future implementation requirements.