language technology
Kabakoo's pilot of a Bamanankan AI mentor achieved a Word Error Rate of 0.64, down from a baseline of 0.98, using just 10 hours of audio data
Kabakoo's pilot of a Bamanankan AI mentor achieved a Word Error Rate of 0.64, down from a baseline of 0.98, using just 10 hours of audio data.
Primary evidence
Kabakoo's pilot of a Bamanankan AI mentor achieved a Word Error Rate of 0.64, down from a baseline of 0.98, using just 10 hours of audio data.
Kabakoo reduced its Bambara AI mentor's Word Error Rate from 0.98 to 0.64 using OpenAI's Whisper architecture, but faces a data gap, having only 10 hours of high-quality Bambara audio data and needing an additional 90 hours to build a scalable model.
Supporting evidence
Across 11,171 bilingual conversations with Kabakoo's AI Mentor, nearly 40% of voice interactions happened in Bamanankan.
Kabakoo developed an AI mentor in Bambara, one of West Africa's major languages spoken in Burkina Faso, Ivory Coast, Mali, and Guinea, to make quality upskilling accessible even for those with low French literacy.
Kabakoo AI System: Transforming Digital Learning in West Africa.