Community-Sourced AI Models for African Language Diversity

Paper: Towards Model as a Library: Offline, Community-Sourced AI for Low-Resource African Languages

Authors: Fendji K. E. Jean Louis

Executive Summary

Recent research highlights significant challenges in using large language models (LLMs) for AI-driven services in African communities. Despite the potential of these models, they struggle with representing the diverse dialects and regional variations of African languages, which are often categorized as low-resource. This inconsistency is particularly concerning because it affects the communities that could benefit the most from such technologies.

The study introduces a new software architecture called Model as a Library (MaaL). This innovative approach allows for the development of small, community-specific speech models that can be used offline. MaaL addresses the issue of unreliable language representation by enabling structured data collection that is grounded in the actual speech patterns of local populations. Instead of relying on large datasets scraped from the internet, which often fail to capture local nuances, MaaL's vocabulary is built from a handful of recordings made by community members. This method ensures that the models better reflect the way people actually communicate.

One of the core features of MaaL is its keyword spotting capability, which transforms a closed vocabulary from text into voice. This allows for effective communication without needing extensive language resources. The system operates entirely on-device, making it particularly suitable for areas with limited internet access and low literacy rates. This offline capability could significantly enhance the reach and usability of digital tools that already serve these communities.

The authors propose that existing digital tools, which often contain closed-vocabulary elements, could be adapted to work within the MaaL framework. This could provide a straightforward way to collect voice data from populations with low literacy, without requiring extensive training or resources.

The research outlines the conceptual and mechanical aspects of MaaL, along with an analysis of its feasibility. It identifies the requirements for implementing a working version of the system, emphasizing the potential impact on underserved communities. By focusing on localized speech models, this approach could facilitate more effective AI applications in African languages, ultimately leading to better service delivery in education, healthcare, and other essential sectors.

Academic Abstract

Large language models are frequently proposed as a route to AI-powered services for African communities, but they are least reliable exactly where the need is greatest: all African languages remain low-resource by any standard measure, and models trained on scraped, standardised text systematically misrepresent the dialectal and regional variation of how people actually speak. We introduce \textbf{Model as a Library (MaaL)}, a software architecture that packages small, community-enrolled speech models as versioned on-device dependencies, enabling offline structured data collection that cannot generatively hallucinate, for populations that current language models serve worst. Rather than relying on web-scraped corpora, MaaL's vocabulary is enrolled directly from a small number of example recordings by the speakers themselves, at the point of deployment. We describe the architecture and its central mechanism - keyword spotting that turns a closed-vocabulary text form into a voice form, filled and submitted entirely on-device - and propose transpiling the closed-vocabulary elements already present in widely-deployed digital form tools into MaaL schemas, a low-friction path to voice-first, offline data collection for the low-literacy populations these tools already reach. This is a position and system-design paper: we describe the concept, the mechanism, and an analytical feasibility case, and identify what a working implementation still requires.