Measuring what is missing, then closing it

The problem

Standard metrics like Word Error Rate treat two utterances as identical when their letters match — even when pitch and tone change the meaning. In tonal and dialectal languages, that difference is the meaning.

Our approach

We design reward signals and human-feedback protocols that capture pitch contour and dialectal nuance, so a model is rewarded for how a sentence is actually said, not merely how it is spelled.

The problem

There is almost no openly-licensed, validated speech data for Sudanese Arabic, Fur, or Zaghawa — and what exists rarely reflects dialectal variation or displacement contexts.

Our approach

Voice-first collection with native speakers and displaced communities, and a human-in-the-loop consensus pipeline for gold-standard labels. Datasets are held internally; how they may ever be shared is set out in Data Rights & Consent.

The problem

Safety frameworks are usually written elsewhere and imported — encoding foreign assumptions into systems that mediate education, health, and finance for African users.

Our approach

We ground our Constitutional AI judge in the African Declaration on Internet Rights and Freedoms (AfDec) and local cultural context, so safety reflects the communities the data serves.

The problem

Many speakers live with intermittent connectivity or none — and many cannot write their language. Standard pipelines assume bandwidth and literacy that simply are not there.

Our approach

Offline-first contribution, parameter-efficient methods, and voice-first interfaces — so being represented never depends on a fast connection or the ability to type.