ASR training data for African languages, starting with Nigeria.
Transcribed speech in Nigerian Pidgin, Yorùbá, Igbo, Hausa and Nigerian-accented English, read and conversational, with code-switching kept and speaker metadata on every clip.
Free 10-hour sample first. A person replies within one working day.
Footage that suits this use case.
Where the data runs short.
Little transcribed audio
Many African languages have far less transcribed speech than English, and much of what exists is read text, not conversation.
Code-switching
Real users mix languages within a sentence. Models trained on clean single-language audio stumble on it.
Licences and consent
Some open sets restrict commercial use. Product teams need clear licences and consent records.
What you get from TrustAm Data.
- Languages
- Nigerian Pidgin, Yorùbá, Igbo, Hausa, Nigerian English; others scoped per order
- Tasks
- Read prompts, natural conversation, service dialogues
- Transcripts
- UTF-8, orthography agreed up front, code-switching tagged on request
- Metadata
- Language, age band, gender, region, device, consent ID and version
- Licence
- Commercial licence summary at /data/legal/licence
Scope it so the first sample is useful.
- Pick languages and the read/conversation split
- Set a speaker quota (region, age band, gender)
- Agree the transcription guide, including tone marks and code-switching
- Choose a test set to hold out by speaker
Provenance on every minute.
ID-checked recorders
BVN or NIN plus a live selfie matched to the ID, and a bank account in the same name, before any paid task.
Consenting workplaces
Each workplace is registered in person by a TrustAm field agent, with the owner’s consent and a location fence.
Live capture only
A spoken challenge code, motion sensors and GPS show each clip was filmed live, at work. No camera-roll uploads.
Blurred and reviewed
Bystander faces and number plates are blurred; a trained reviewer approves or cuts every minute before delivery.
Buyers usually ask.
Should we start with open datasets?
Often, yes. Open sets such as Mozilla Common Voice and academic corpora are good for a first model and for benchmarking. Teams come to us when they need more hours, conversation, code-switching or a clean commercial licence.
Can you hold out speakers for a test set?
Yes. Every speaker has a stable pseudonymous ID, so we can split by speaker and deliver train and test sets that never share a voice.
How do I start?
Email [email protected] or use the request form. A person replies within one working day and we record a free sample of about 10 hours to your spec. Pricing is quoted per approved hour after scoping.
Tell us what you need. We send a free sample first.
Email [email protected] with the data type, hours, languages or trades, labels and format. Pricing is quoted per approved hour after scoping.