Large language models are often touted as a way to deliver AI services to African communities, yet they perform worst on the languages that need them most. All African languages are low‑resource by any metric, and models trained on scraped, standardized text systematically miss the dialectal and regional variation of real speech.
We introduce 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. MaaL processes everything locally, allowing the most underserved populations to benefit from reliable AI.
The central mechanism is keyword spotting, which turns a closed‑vocabulary text form into a voice form that is filled and submitted entirely on‑device. The vocabulary is enrolled directly from a handful of recordings provided by speakers at deployment time, rather than from web‑scraped corpora.
We propose transpiling the closed‑vocabulary elements already present in widely‑deployed digital form tools into MaaL schemas, offering a low‑friction, voice‑first, offline data‑collection path for low‑literacy users.
This paper is a position and system‑design article: it describes the concept, the mechanism, an analytical feasibility case, and outlines the remaining requirements for a working implementation.
Review