Source-linked AI summary
Towards Model as a Library: Offline, Community-Sourced AI for Low-Resource African Languages
Fendji K. E. Jean Louis
TL;DR
African-language communities face unreliable model performance associated with low-resource data and barriers from unreliable connectivity. MaaL proposes small, community-enrolled speech models packaged for on-device, offline structured data collection. Its analytical feasibility estimates cover storage and latency, while field accuracy and other implementation questions remain open.
Problem
African languages remain low-resource, and models trained on standardised text can misrepresent dialectal and regional variation; unreliable connectivity can also block access to language tools.
Source: §1 · pp. 1–2Method
Results
The paper reports analytical storage and latency estimates, not deployed-system measurements, and does not address recognition accuracy under real acoustic conditions.
Source: §5 · p. 4Takeaways & Limitations
The paper proposes reusing closed-vocabulary elements in existing digital forms as a route to MaaL schemas without discarding current infrastructure.
Source: §4.1 · p. 3, §7 · p. 5
Abstract
from arXiv · showhide
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.
1 Introduction
The introduction argues that data scarcity and connectivity barriers make current language-model services least reliable for many African-language communities. It proposes MaaL as an offline, community-enrolled alternative for structured data collection, with a path to reuse existing forms.
1 Introduction
- On WARRI, models score 76.3–83.4 ChrF++ on standardised BBC-register Pidgin but approximately 54 on community-level Naija, a 24.3-point gap linked to linguistic guesswork outside training distributions [1] [5].The paper attributes the gap to training data that overrepresents one register and erases others.Source: §1 · p. 1
- Low-resource levels directly correspond to lower-quality, less reliable models, while unreliable connectivity and limited linguistic coverage can jointly block access.The paper describes a live connection to a distant server as a second barrier for this population.Source: §1 · pp. 1–2
- MaaL packages small keyword-spotting models as versioned on-device dependencies, with vocabulary enrolled from a few speaker recordings and no cloud round-trip or retraining.Closed-set inference returns a community-provided term or an explicit rejection rather than a generated response.Source: §1 · p. 2
- The paper proposes transpiling closed-vocabulary elements in ODK and KoboToolbox forms into MaaL schemas, offering a route to unassisted forms without new data-collection infrastructure.The paper presents this as a contribution, alongside an agricultural-survey scenario and analytical feasibility estimates.Source: §1 · p. 2
2 Related Work
Related work positions MaaL as packaging existing small-model and keyword-spotting techniques for offline data collection, rather than proposing a new matching method.
2 Related Work
- MaaL packages existing few-shot and query-by-example keyword-spotting methods as versioned dependencies rather than proposing a new method.Prototypical approaches enroll keywords from a handful of examples, open-set variants distinguish enrolled from out-of-vocabulary input [9], and on-device adaptation has been demonstrated under sub-10KB RAM budgets [3].Source: §2 · p. 2
- InkubaLM, a 400M-parameter model trained on curated data across five African languages, matches or exceeds larger models on in-language tasks [11].The paper presents it as evidence supporting deliberately sourced small models for underrepresented languages.Source: §2 · p. 2
- Existing edge platforms require cloud retraining and a full firmware reflash for vocabulary changes [7], while ODK and CommCare support offline form synchronisation but leave AI inference cloud-side.MaaL is positioned as an offline inference layer for such data-collection platforms.Source: §2 · p. 2
- Cloud-API access to large models is described as structurally unsuited to intermittent-connectivity deployment [12].Source: §2 · p. 2
3 The MaaL Architecture
MaaL uses closed-set nearest-prototype matching to map speech to enrolled labels or reject it, with a form schema, on-device runtime, and shared-backbone model library. Closed-set output avoids generative fabrication but does not remove consequential classification errors.
3 The MaaL Architecture
- A MaaL model embeds input and compares it with enrolled keyword prototypes by cosine similarity, accepting the best match only when it clears confidence threshold τ.Its tuple comprises a frozen embedding backbone, prototype store, label map, threshold, and input modality.Source: §3 · p. 2
- Closed-set matching cannot invent an unenrolled term, but a wrong enrolled term can be falsely accepted and a correct term falsely rejected.These classification errors can cause real harm in health and agricultural data collection and are not analysed here.Source: §3 · pp. 2–3
- The architecture has a form schema, an entirely on-device runtime, and a model library combining a shared embedding backbone with independently versioned domain models.The schema declares the model for each field; the runtime prompts, records, routes, and applies the threshold. Each domain model contributes roughly 512 bytes per enrolled term.Source: §3 · p. 3, Table 1 · p. 3
- New terms are enrolled by passing a few example recordings through the frozen backbone and storing their mean embedding, without gradients, a GPU, or connectivity.This allows field workers to add locally specific terms in minutes.Source: §3 · p. 3
4 Illustration: From a Text Form to a Voice Form
The illustration maps a structured agricultural survey to reusable domain-model types and a speaker-verified vocabulary. It then shows how closed-vocabulary fields in existing forms could be transpiled into MaaL schemas while retaining form logic and enabling offline, unassisted completion.
4 Illustration: From a Text Form to a Voice Form
- The survey illustration uses agri_term, numeric, time_period, and yes_no models for crop names, quantities, seasons, and input-use responses.These types are presented as covering a large share of typical structured-survey fields and as reusable across forms and deployment languages.Source: §4 · p. 3
- Numeric fields remain an unresolved design question because spoken quantities are compositional and nearest-prototype matching alone does not explain arbitrary multi-digit numbers.Discretised ranges or digit-level composition are proposed possibilities, not specified solutions.Source: §4 · p. 3
- For each deployment language, vocabulary is drawn from community-verifiable sources and confirmed with speakers, rather than scraped from web text.Unconfirmed terms are to be marked for verification rather than guessed.Source: §4 · p. 3
- At runtime, an accepted spoken match writes a normalised value into its form field, and the completed record is submitted as a text-form submission without reading or typing fields.Source: §4 · p. 3
4.1 From Text Forms to Audio-First Forms
- Existing digital-form interfaces assume a respondent or intermediary can read, while closed-vocabulary fields already contain the option lists and answer values MaaL schemas need.The paper names ODK, KoboToolbox, and similar XForm-based platforms as tools in active use across health, agriculture, and humanitarian programmes.Source: §4.1 · p. 3
- The proposal mechanically maps closed-vocabulary fields to domain models seeded with existing options and community-recorded prototypes, without changing form logic or question order.Source: §4.1 · pp. 3–4
- MaaL is framed as an accessibility layer for existing infrastructure, allowing respondents to complete forms unassisted, in their own language, and offline.Building and evaluating the transpiler is identified as a near-term extension.Source: §4.1 · p. 4
5 Feasibility
MaaL’s analytical feasibility estimates suggest its storage and latency fit commodity Android and microcontroller-class budgets, but they are not measurements of a deployed system and do not establish recognition accuracy.
5 Feasibility
- ∼46–52 KB is the cited footprint range for the smaller INT8 DS-CNN variant; ∼190 KB for DS-CNN-M is an extrapolated upper bound.The two smaller-variant measurements are 52.5 KB and 46.15 KB; the larger estimate is not a direct measurement.Source: §5 · p. 4, Table 2 · p. 4
- The resulting footprint and latency estimates fit commodity Android and microcontroller-class hardware budgets and the natural pacing of spoken interaction.Source: §5 · p. 4
- The estimates cover storage and latency, not recognition accuracy, which depends on few-shot enrolment quality under real acoustic conditions.The paper identifies implementation and evaluation on field-collected audio in a target language as its central open question.Source: §5 · p. 4
6 Discussion and Open Questions
MaaL deliberately limits itself to enrolled vocabulary and identifies field accuracy, language coverage, schema robustness, and community governance as unresolved questions.
6 Discussion and Open Questions
- MaaL intentionally excludes open-ended and out-of-vocabulary requests; richer needs call for explicit escalation to a community-trained generative model such as InkubaLM [11].Source: §6 · p. 4
- Open questions include field-acoustic accuracy and misrecognition harms, shared backbones across regional languages, transpiler robustness, and governance of community vocabulary.Source: §6 · p. 4
7 Conclusion
MaaL proposes small, community-enrolled closed-set models as an offline alternative to language models that are least reliable where linguistic need is greatest. Reusing closed-vocabulary elements in existing forms offers an adoption route without replacing that infrastructure.
7 Conclusion
- MaaL proposes small, community-enrolled, closed-set models that work offline and grow vocabulary from language as spoken rather than scraped corpora.Source: §7 · pp. 4–5
- MaaL targets a data gap that scale alone cannot fix: large language models are least reliable where linguistic need is greatest.Source: §7 · p. 4
- Existing digital forms already contain closed-vocabulary structures that can supply MaaL schemas without requiring organisations to discard their infrastructure.Source: §7 · p. 5