A working shelf
A low-resource pair is rarely “no data”. It is data in the wrong place, the wrong licence, the wrong register, or a Drive folder someone emailed in 2018. This shelf is opinionated toward West and Central Africa because that is where we have spent the most time. It is not exhaustive.
For where languages are spoken, languageplayer.io and CLEAR / TWB’s language-use data are the right first maps. Maps do not train models. They stop you training the wrong one.
Speech
ALFFA (African Languages in the Field: speech Fundamentals and Automation), OpenSLR 25, remains the first ASR baseline many Wolof papers report against: field-recorded African speech, not studio English. BibleTTS adds Asante Twi, Akuapem Twi, Ewe, Hausa, Lingala, and Yoruba under CC BY-SA 4.0. The recordings are excellent and liturgical. They are dangerous if they are all you train on.
On Wolof, three checkpoints are worth naming together: Oolel, SpeechBrain’s wav2vec2 (ALFFA WER 16.25), and abdouaziiz’s XLS-R (ALFFA_PUBLIC WER around 0.21). WER on ALFFA is not WER on a noisy motorbike voice note. Treat the numbers as ordering.
SIL Bloom speech fills languages absent from Common Voice: Bambara, Soninke, Mamara Sénoufo. RobotsMali’s Jeli ASR and a 31-hour Bambara set are what a national-lab speech programme looks like when it is real. Senelangues holds light archives for Serer, Ménik, and neighbours — small, precious, easy to overfit. Dioula has a public lever in Common Voice.
Dictionaries and bitext
Neural translation without a lexicon is how you invent a word for “mosquito net” that no nurse uses. DiLAF still matters: French to Bambara, Hausa, Kanuri, Nouchi, Tamajaq, Wolof, Zarma. Kanuri is a Sahel language of Niger, Nigeria, Chad, Cameroon, and into Sudan. Nouchi is Abidjan. Tamajaq is Tuareg. Zarma is Songhai of Niger. A bilingual dictionary is a constrained parallel corpus with unusually high precision.
For question-shaped text, Masakhane’s AfriQA queries cover Bembe, Fon, Hausa, Igbo, Kinyarwanda, Swahili, Twi, Wolof, Yoruba, and Zulu — a different register from news, useful in evaluation even when you do not train on it. The community dataset list should be read before anyone proposes a new crawl. Lacuna Fund has financed sets for Hausa, Igbo, Nigerian Pidgin, and Yorùbá; for Afar, Amharic, Oromo, Somali, and Tigrinya; for Kiswahili and the Dholuo / Luhya cluster; and a twenty-language pack stretching from Bambara to isiZulu. Funded data still has licences.
A few languages we keep coming back to
Google Translate already “covers” Bambara; Google even told the story. Coverage is not a Malian radio workflow. RobotsMali’s Bayelemabaga (~45k lines) sits beside Kumatigi and the audio above. Even a lyrics page such as Mandjou is a reminder that register is not a news crawl. Fine-tuning Bambara is no longer a race to be first in the language. It is a question of whose system, and for which use.
Nande, a Bantu language of central Africa, has a named researcher, a French pair, a published checkpoint, and a countable corpus: nnd_fr_mt_v3 and nnd_fr_26k. It is not waiting on a consumer catalogue.
Wolof is the rare case where speech papers are ahead of the average translation blog post. Scripture is on eBible; dofbi/jolof is on Hugging Face. Qur’an bitext circulates on unstable Drive links — check the licence before training.
Serer and Fula share noun-class suffixes. That is a linguistic fact, not a modelling trick: if you are pivoting or sharing subword vocabulary between Seereer and Pulaar, you have a reason. Do not assume mutual intelligibility; morphology will leak in useful ways if you are careful. Soninke has a Bloom speech model; Soussou is still thinner in public ASR; Dioula has Common Voice. Mande languages are a family. They are not a single training file.
Licence first. Register mixed on purpose. Speech is bitext you have not typed yet. Cite the lab that already did the collection — RobotsMali, Masakhane, SIL, Salomon Metre, Soynade. If you have a corpus that should be on this shelf, write to hello@tokyma.io. We would rather link than duplicate.