The Problem

The overwhelming majority of the world's languages have no meaningful representation in current AI systems. State-of-the-art large language models are cloud-centralised, computationally heavy, and dependent on reliable connectivity, making them ill-suited to the very environments where many endangered and minority languages are actually spoken.

When AI does reach these communities, the prevailing model is extractive: linguistic data is taken from communities, frequently without sustained consent or governance, to train proprietary models that serve outside commercial interests. Communities feed data into systems they do not control.

Kinyarwanda, spoken by over 12 million people, remains poorly served by mainstream AI platforms. This system was built to address that gap directly.

What This System Does

Vuganeza(Ijambo) is a multimodal, field-deployable language-learning application. It integrates translation, speech synthesis, pronunciation evaluation, and spaced-repetition learning into a single system built specifically for Kinyarwanda.

📖

AI Interactive Reader

Input any text in English or Kinyarwanda. The system translates sentence-by-sentence with word-by-word mapping. Words start hidden. The learner taps to reveal, creating an active recall moment. Every tapped word is automatically captured for spaced repetition review.

🧠

Spaced Repetition (SM-2)

The SM-2 algorithm, the same engine behind Anki, schedules vocabulary review at scientifically optimal intervals. The telemetry from this system (acquisition speed, retention per study hour, forgetting curves) provides a direct, quantifiable measure of human learning, not model accuracy.

🔊

Kinyarwanda Speech Synthesis

Powered by Meta's MMS-TTS model trained on Kinyarwanda speech data. A dedicated neural model for the language, not a generic text-to-speech engine. Supports Krashen's comprehensible-input hypothesis by delivering spoken language calibrated to the learner's current level.

🎙

Pronunciation Evaluation

Uses a fine-tuned Whisper model (mbazaNLP/Whisper-Small-Kinyarwanda, ~24% WER on Common Voice) to transcribe the learner's spoken Kinyarwanda and score pronunciation via normalised word-level edit distance.

💡

Structural Grammar Breakdown

AI-generated grammatical analysis following the Michel Thomas method: literal blueprint, building blocks (noun classes, verb prefixes), and a generative rule the learner can apply immediately to construct new sentences.

📊

Learning Telemetry

All learning data (words studied, retention rates, pronunciation scores, session frequency) is stored in SQLite and remains under the learner's control. This telemetry is designed to serve as a direct, quantifiable measure of acquisition in field research settings.

Technical Architecture

→

Translation Engine: NLLB-200 (Meta), 600M parameters, optimised through CTranslate2 with int8 quantisation for low-latency inference on CPU

→

Speech Synthesis: MMS-TTS Kinyarwanda (Meta's Massively Multilingual Speech), dedicated Kinyarwanda voice model

→

Speech Recognition: Whisper-Small fine-tuned on Kinyarwanda Common Voice data (mbazaNLP), ~24% WER

→

Learning Engine: SM-2 spaced repetition algorithm with per-word ease factor, interval scheduling, and retention telemetry

→

Grammar Analysis: LLM-powered structural breakdown (DeepSeek V4 Flash) using the Michel Thomas pedagogical method

→

Data Layer: SQLite with WAL mode. All learning data stored locally, never transmitted to external servers without explicit consent

Data Sovereignty by Design

The central theoretical commitment of this project is to decolonial AI. The Indigenous data sovereignty movement, condensed in the CARE principles, holds that communities should govern the collection, holding, and use of their own data. Most AI projects endorse this principle rhetorically but fail to enforce it architecturally.

In this system, sovereignty is not a policy promise. It is an engineering constraint. The primary database sits locally in SQLite. The learner's interaction data, vocabulary progress, pronunciation attempts, and session records never leave the system without explicit, revocable consent. The community, not the researcher and not a cloud provider, decides what happens to that data.

This architecture was validated with Kinyarwanda. The proposed Master's research would test whether it gives a different language community, in Bahia, a genuine sense of control over its linguistic heritage, and what governance challenges arise in practice.

Research Context

This system was built for Kinyarwanda as a working proof of concept. The proposed Master's dissertation, titled "Sovereign and Multimodal AI: Integrating Governance with Field-Deployable Systems for Accelerated Language Acquisition," would adapt and deploy this architecture with a regional language community in Bahia, Brazil, through PPGTEL at Universidade do Estado da Bahia (UNEB), DCH Campus V, Santo Antônio de Jesus.

The research sits within the line Linguagens, contextos e letramentos, and proposes a participatory action research methodology aligned with Indigenous data-governance frameworks and Brazilian research ethics norms.

The study aims to answer two questions: whether the tool accelerates vocabulary acquisition (measured through spaced-repetition telemetry), and whether an on-device data architecture gives a community genuine governance over its own linguistic data. Kinyarwanda is the engineering foundation. Bahia is the proposed research site.

About the Researcher

Adu Emmanuel Ayorinde

Software developer and founder of PhotonCognis Limited, based in Kigali, Rwanda, with roots in Abuja, Nigeria. Specialises in Python, AI integration, and digital product architecture.

Built Vuganeza(Ijambo) as a working proof of concept for community-centred language technology, designed from the ground up for a language that mainstream AI platforms largely ignore.

The proposed Master's at PPGTEL represents the transition from engineering to empirical research: studying, with the tools of applied linguistics and language pedagogy, whether the systems he builds genuinely serve the communities they claim to serve.

Key References

Krashen, S. D. (1985). The Input Hypothesis: Issues and Implications. London: Longman.

Tabibian, B. et al. (2019). Enhancing human learning via spaced repetition optimization. Proceedings of the National Academy of Sciences, 116(10), 3988–3993.

Kwok, J. C. Y. (2026). Data colonialism and indigenous languages in AI. AI and Society, online first.

Mohamed, S.; Png, M. T.; Isaac, W. (2020). Decolonial AI: Decolonial theory as sociotechnical foresight. Philosophy and Technology, 33(4), 659–684.

Lilley, S. et al. (2024). Māori data sovereignty: Contributions to data cultures. Information, Communication and Society, 27(16), 3099–3114.

Tapo, A. A. et al. (2024). Machine intelligence in Africa: a survey. arXiv:2402.02218.

Li, P.; Lan, Y.-J. (2022). Digital Language Learning (DLL). Bilingualism: Language and Cognition, 25(3), 361–378.

What do you want to read?

Pick a text below, or paste your own. The AI will turn it into an interactive lesson.

Vuganeza(Ijambo)

The Problem

The Research Question

What This System Does

AI Interactive Reader

Spaced Repetition (SM-2)

Kinyarwanda Speech Synthesis

Pronunciation Evaluation

Structural Grammar Breakdown

Learning Telemetry

Technical Architecture

Data Sovereignty by Design

Research Context

About the Researcher

Key References

Try the Prototype

Vuganeza(Ijambo)

AI Interactive Reader

What do you want to read?

Beginner Readings

Curated Lessons

Verb Conjugator

Compilation Formula

All caught up!

Review Time

Pronunciation Practice

Your Progress