Model as a Library Proposes Offline Voice Data Collection for Low-Resource African Languages

Summary

Large language models are often proposed for AI services in African communities, yet the paper argues that African languages remain poorly served and that scraped, standardized text can misrepresent dialectal and regional speech variation. It introduces Model as a Library (MaaL), a software architecture that packages small speech models enrolled by local communities as versioned, on-device dependencies. Speakers provide a small number of example recordings directly at deployment, allowing the vocabulary to reflect local usage rather than relying on web-scraped corpora. MaaL uses keyword spotting to convert closed-vocabulary digital forms into voice forms that can be completed and submitted entirely offline, without generative text generation that could hallucinate. The authors also propose transpiling closed-vocabulary elements from widely deployed digital form tools into MaaL schemas, providing a path to voice-first data collection for low-literacy populations already reached by those tools. The work is a position and system-design paper with an analytical feasibility case; it does not report a completed implementation and identifies implementation requirements that remain unresolved.