AI Resources

Nigeria's first government-backed multilingual LLM (Sep 2025): a Llama-3 8B fine-tuned on 400M+ tokens across 4 Nigerian languages. Produced by NCAIR/NITDA and Awarri.

Open-source fraud-intelligence dataset for Nigerian fintechs (Mar 2026), the first of its kind. Relevant to payment fraud, KYC fraud and account-takeover detection.

A massively multilingual masked language model covering 517 African languages and varieties across five scripts, achieving state-of-the-art results on the AfroNLU benchmark. Developed by the UBC Deep Learning and NLP Lab as an Afrocentric resource.

Spitch is a Lagos-based voice-AI company (founded Oct 2024) providing ASR (speech-to-text), TTS (text-to-speech) and translation APIs/SDKs for Nigerian languages (Yoruba, Igbo, Hausa, Nigerian-accented English, plus Amharic) so teams can add local-language voice to call centres, media and learning tools.

SALT (Sunbird African Language Technology) is a multi-way parallel text-and-speech corpus of English plus Luganda, Acholi, Lugbara, Ateso, Runyankole and Swahili supporting ASR, TTS and translation, with companion Whisper and MMS models. Built with Makerere University AI Lab.

An Afrocentric many-to-many machine translation model (1.2B params, mT5) fine-tuned from Cheetah to support 156 African language pairs, evaluated on the AfroLingu-MT benchmark. Developed by the UBC Deep Learning and NLP Lab and published at ACL 2024.

UlizaLlama is a 7B-parameter Swahili-and-English LLM fine-tuned from Meta's Llama 2 (continually pretrained on ~321M Swahili tokens) by Jacaranda Health in Kenya, built to power Swahili maternal-health SMS support for low-income expectant mothers in East Africa.

Vulavula is Lelapa AI's API platform for speech-to-text transcription and machine translation across South African and African languages (isiZulu, Sesotho, Sepedi, Setswana, isiXhosa, Afrikaans, English, Swahili, Hausa, Yoruba, French), purpose-built for telco and financial-services contact centres.

Africa's largest data science and AI competition platform where organizations host real-world ML challenges and a community of builders competes to solve them. Offers competitions, learning courses, jobs and leaderboards, with partners including Microsoft, Google, AWS and Google DeepMind.

AfriCOMET is a COMET-based machine translation evaluation model for African languages, scoring translation quality from source, hypothesis and reference triplets. The STL-1.1 version uses the afro-xlmr-large-76L encoder and was validated in the WMT 2024 Metrics Shared Task across 13 African-centric language pairs. It is released by the Masakhane community with reference-based and quality-estimation variants.

AfriHuBERT is a compact self-supervised speech representation model based on mHuBERT-147, continually pretrained via multilingual adaptive finetuning on over 10,000 hours of speech spanning more than 1,200 African languages and varieties. It improves spoken language identification and ASR over its base model and acts as an encoder for downstream African speech tasks. Its training data was aggregated from sources including BibleTTS, Kallaama, NaijaVoices and NCHLT.

AfriMT5 (afri-mt5-base) is an mT5-based machine translation model from the Masakhane community fine-tuned to translate across 16 African languages using multilingual adaptive fine-tuning. It targets news-domain translation for low-resource African languages, several of which were not previously covered by existing benchmarks. It is distributed as open weights on HuggingFace.

[ingest-scout] Suite of 10 open large language models from McGill-NLP adapted for 20–50 African languages via continued pre-training on 26–35B tokens; includes AfriqueLlama-8B, AfriqueQwen variants (4B–14B), and AfriqueGemma variants (4B–12B). Backed by ACL 2026 research paper (arXiv:2601.06395); models updated June 2026 with active downloads. Covers languages not served by registry entries (Amharic, Tigrinya, Oromo, Wolof, Bambara, Moroccan/Egyptian/Tunisian Arabic, Malagasy, Afrikaans, etc.).

Imported from a community submission, needs review.

Bambara-ASR-v2 is an automatic speech recognition model for Bambara (Bamanankan) fine-tuned from OpenAI's Whisper-large-v2 using parameter-efficient tuning, reaching about 25 percent word error rate. It handles natural Bambara-French code-switching common in Mali's multilingual context. It is released under Apache 2.0 as part of the MALIBA-AI community initiative.

[ingest-scout] Rwanda-based NLP organization Digital Umuganda's Conformer-CTC ASR model trained on 660 hours of Kinyarwanda speech (Afrivoice dataset); last updated December 2 2025, 52 monthly downloads, CC-BY-4.0 license, built on NVIDIA NeMo. Distinct from the already-inventoried mbaza-whisper-small-kinyarwanda — this is a larger NeMo-based model; fills Rwanda AI infrastructure representation.

Imported from a community submission, needs review.

[ingest-scout] Ethiopian language ASR model fine-tuned from Facebook w2v-bert-2.0 (600M params) on Amharic speech, achieving 22.37 WER on the WAXAL Amharic test set; paper published March 2026 (arXiv:2603.23654), dataset updated June 2026, 1.03k monthly downloads, CC-BY-4.0. Developed by the Ethio-ASR research team — a verified live model filling the Ethiopia AI gap in the registry.

Imported from a community submission, needs review.

EthioLLM is a family of multilingual language models (XLM-RoBERTa and mT5 based) for five Ethiopian languages: Amharic, Ge'ez, Afaan Oromoo, Somali and Tigrinya, plus English. The large variant EthioLLM-l-70K is a fine-tuned XLM-RoBERTa-Large used for masked language modeling and downstream tasks like classification, NER and sentiment. It was released by the EthioNLP collective alongside the Ethiobenchmark evaluation suite.

xTTS-v2-wolof is a text-to-speech model for Wolof built by GalsenAI by fine-tuning Coqui's xTTS v2 on the cleaned Anta Women Wolof TTS dataset. It synthesizes Wolof speech and can clone a voice from a reference clip as short as 6 seconds. It was developed by the GalsenAI community in Senegal.

IrokoBench is a human-translated evaluation benchmark for 16 typologically diverse African languages covering three tasks: natural language inference (AfriXNLI), knowledge QA (AfriMMLU) and mathematical reasoning (AfriMGSM). It is widely used to measure the performance gap between English and African languages in large language models. It was released by the Masakhane community and published at NAACL 2025.

Lugha-Llama is a Llama-3.1-8B model continually pretrained on the WURA African-language corpus to lift performance on low-resource African languages. It ships in three variants (wura, wura_edu, wura_math) and reaches leading results among similarly sized models on the IrokoBench and AfriQA African-language benchmarks. It was built by researchers at Princeton University.

MALIBA-AI Bambara TTS is a neural text-to-speech model for Bambara (Bamanankan), the most widely spoken language in Mali, built on the Spark-TTS framework with a Qwen2.5-based backbone of around 500M parameters. It supports 10 authentic Bambara speaker voices and outputs 16kHz mono audio without a separate vocoder. It is released under a non-commercial MALIBA-AI research license.

Whisper-Small-Kinyarwanda is an automatic speech recognition model fine-tuned from OpenAI's Whisper-small on the Common Voice Kinyarwanda dataset, achieving about 24 percent word error rate. It transcribes Kinyarwanda audio into text for speech applications. It was built by Mbaza NLP, part of the Digital Umuganda voice-technology ecosystem in Rwanda.

NaijaVoices is a large-scale speech dataset of about 1,800 hours from over 5,000 speakers with expert-curated transcripts in Igbo, Hausa and Yoruba, roughly 600 hours per language. It is designed for building ASR and speech AI for Nigerian languages and improves Whisper and MMS fine-tuning performance. It is available on HuggingFace behind a free registration and powers models like AfriHuBERT and SBPN.

NileChat-3B is a 3B-parameter LLM built on Qwen2.5-3B and adapted for Egyptian and Moroccan communities, handling Egyptian and Moroccan dialectal Arabic in both Arabic script and Arabizi alongside Modern Standard Arabic, French and English. It was trained with controlled synthetic data and translated corpora and outperforms its Qwen2.5-3B baseline on Arabic benchmarks. It was released by the UBC Deep Learning and NLP Lab and accepted at EMNLP 2025.