Data
Hausa as people speak it, recorded with their consent
Speech models learn the language they are given. We are building the Hausa and Zarma speech that public datasets leave out, with the people who speak it.
Problem
Public Hausa speech is narrow
Most public Hausa speech data was read aloud from a script, in Nigeria, by people comfortable with a phone app. FLEURS, a benchmark many papers report, has a Hausa test set labelled Nigerian Hausa in which 620 of 621 utterances are spoken by women. That is not the Hausa our users speak, and models trained and tested on it learn a narrow version of the language.
Our approach
Collect it together, with clear rules
Rumbu is our programme to build the Hausa speech that is missing, the way a community keeps a granary: together, with care, and with clear rules about who may draw from it. We will record Hausa from Niger and Nigeria in its regional varieties: conversation between two people on separate channels, speech over ordinary phone calls as well as apps, and people who do not read, prompted by pictures and questions rather than text.
Every speaker will give informed consent in Hausa, choose what their voice may be used for, and be paid for their time. Each dataset will ship with a data statement that says who is in it, how it was recorded and what it cannot support.
We will release what consent and law allow, and say plainly what we keep closed and why.
What our data will includeRumbu is opening: nothing has been collected yet
Two-person conversation
Speakers on separate channels, so overlap, backchannels and interruptions survive for full-duplex research.
Ordinary phone calls
Speech over the phone lines people actually use, as well as apps.
Regional varieties
Hausa from Niger and Nigeria, written in the orthography each country uses.
Frozen test sets
Speaker-disjoint, dialect- and gender-stratified, frozen before training data is collected.