Researchers are pitching a way to collect African-language voice data without the internet, by skipping generative AI entirely.
A new arXiv position paper calls the system Model as a Library, or MaaL. It packages small speech models enrolled from a handful of recordings made by a community's own speakers, not scraped from the web. Those models run on-device and use keyword spotting to turn a closed, pre-defined vocabulary into a voice interface, letting someone fill out a form by speaking instead of typing. The authors also propose converting the closed-vocabulary fields already built into existing digital form tools (the kind used for surveys and data collection in the field) into MaaL schemas, so the voice layer piggybacks on software that already reaches low-literacy users.
Every African language counts as low-resource by the usual measures, and models trained on generic scraped text flatten the dialect and regional variation that actually shows up in speech. MaaL's bet is that a closed vocabulary cannot generatively hallucinate: it can only match what it was enrolled to recognize, which matters more for structured data collection than for open-ended chat. That is a narrower ambition than AI that speaks your language, but arguably a more honest one.
Worth flagging: the authors call this a position and system-design paper, not a working product. The feasibility case is analytical, and they say plainly what an actual implementation still requires.