Buying AI Training Data From Nigeria: A Buyer's Guide

Play a speech model a Lagos market conversation: "Abeg, how much you go sell this one for me?" Then watch the transcript. Most systems trained on American and European voices stumble on Nigerian Pidgin, on Yoruba tones and on Nigerian-accented English. Robot models have the same problem with hands: they have seen far more kitchens in a few rich countries than workshops in Aba or Ladipo. This guide is for teams who want to fix that by buying data from Nigeria.

Why source AI training data from Nigeria?

Because the people your model will meet are missing from the data it learned from. Nigeria is one of the largest countries in the world by population, its people speak hundreds of languages, and a large share of its work is done by hand in open workshops and markets. Very little of that has been recorded in a form AI can use.

• Speech gaps. The AfriSpeech-200 authors released 200 hours of African-accented English from 2,463 speakers and still described speech data as non-existent for most African accents.
• Video gaps. Ego4D, the best-known first-person dataset, was filmed in 9 countries. Skilled trades as practised in West Africa are barely there.
• National priority. Nigeria's National AI Strategy lists poor access to quality, standardised data as a barrier to building AI in the country.

What data can you buy from Nigeria today?

TrustAm Data collects two kinds of data in a Lagos pilot: first-person video of skilled work and Nigerian speech, both from ID-verified contributors. Browse what is in stock or collectable to order in the dataset library.

Buying Nigerian speech data: Pidgin, Yoruba, Igbo and Hausa

Specify the language, the region and the kind of speech you need, because "Nigerian speech" is many different things. A clear spec saves you paying for hours that do not help your model.

1. Tone and diacritics. Yoruba and Igbo are tonal, and meaning can change with tone. Decide whether transcripts must carry tone marks.
2. Nigerian Pidgin. There is no single fixed spelling. Agree a spelling guide before transcription starts.
3. Code-switching. Many Nigerians move between English and a Nigerian language in one sentence. If your users do, your data should too.
4. Conversation vs prompts. Natural conversation is messier and more valuable for real-world recognition; prompted speech is easier to control.
5. Region and dialect. Ask for region and dialect in the metadata. Community projects such as Masakhane and Mozilla Common Voice are useful for seeing what already exists in open data.

Consent, licensing and the NDPA: what to ask a vendor

Ask to see the consent record behind any clip in your sample. Data about people in Nigeria falls under the Nigeria Data Protection Act 2023, which the Nigeria Data Protection Commission enforces; its GAID 2025 directive sets out how consent works and says withdrawing it must be as easy as giving it. If you serve customers in Europe, Article 10 of the EU AI Act also expects documented data governance for training data of high-risk systems.

• Consent per person: on TrustAm Data each contributor says the consent line on video in their own language, and the version and date ship in your manifest.
• Consent per place: the owner of every workplace consents before the first recording there.
• Licence scope: who may use the data, for training which kinds of systems, in which territories, and for how long.
• Withdrawal: what happens when a contributor withdraws. We remove their clips from future deliveries and ask buyers who already received them to delete them.
• Bystanders: faces of other people are blurred before delivery.
• Documentation: a datasheet that explains how the data was collected, in the spirit of Datasheets for Datasets.

The trust and consent page covers the full process.

How to judge the quality of a data vendor

Quality comes from who records, how fraud is caught and how much a person actually reviews. Ask every vendor the same questions and compare the answers.

1. Is every contributor a real, different person? Ours pass a BVN or NIN check with a live selfie, are paid only to an account in the same name, and duplicate IDs and faces are refused.
2. Can they prove a clip was filmed live, where they say? Ours start with a spoken 3-digit challenge code with both hands shown, and carry motion-sensor and GPS data inside a 150 m workplace fence.
3. What is checked automatically? Black frames, frozen video, duplicates across the whole collection and phones lying still.
4. Does a person review every minute? Ours do, and contributors are paid only for approved minutes, which keeps incentives aligned with yours.
5. What does the QA report show? Approval rates and cut reasons for every batch, so you can see what was removed and why.

Delivery formats for video and speech datasets

Ask for formats that drop straight into your pipeline. Our standard delivery is a signed download link, valid for seven days, containing:

• Blurred, trimmed MP4 video (H.264, 1080p at 30 fps).
• Per-clip JSON metadata: trade, duration, capture date, device model, motion-sensor trace and QA score.
• A manifest file with clip ID, pseudonymous contributor ID, workplace ID, consent ID and consent version.
• A QA report with approval rates and cut reasons.
• LeRobot or RLDS format on request, for robot-learning pipelines.

From a free sample to a pilot to monthly supply

There is no self-serve checkout. A person on our team scopes each order with you, and you see real data before you pay anything.

1. Free sample. Tell us the activity, trades, languages and labels. We record about 10 hours to your spec, review every minute and send it with a QA report.
2. Paid pilot. Between 100 and 500 hours, priced per approved hour and quoted after scoping. We open a batch of verified contributors for your order.
3. Weekly batches. Approved minutes arrive every week with the QA report.
4. Monthly contract. When the pilot works for you, we agree a monthly volume, label layers and delivery format, and keep collecting.

More detail: TrustAm Data for AI and robotics labs and custom data collection. If you want the technical detail on capture and fraud controls, read Egocentric video data collection in Africa.

How to request a sample

Email [email protected] or use the request form. A person on our team reads every request and replies within one working day. To get an accurate scope, include:

• The activity or task you need (for example, "hand-sewing and machine-sewing, first-person").
• Trades and languages, and any region or dialect needs.
• How many hours, and by when.
• Label layers and delivery format (MP4 plus JSON, LeRobot, RLDS).
• Your licence needs: use, territory and term.

Are you a Nigerian who wants to record rather than buy? Read How to earn money from AI data in Nigeria or go to trustamai.com/earn.