Skip to content
Kabakoo AcademiesEvidence

Evidence Cases

Bamanankan speech recognition

Among surveyed Kabakoo learners, 92% have difficulty reading Bamanankan and half can't read it at all, while learners consistently prefer oral Bamanankan.

Starting with Bamanankan

In the AI Mentor's conversation data, text stays overwhelmingly French while nearly 40% of voice interactions happen in Bambara. Text-based AI in Bamanankan, which the AI Mentor already does, only covers part of the need. Voice is the real frontier.

How we built the dataset

We launched our AI coach on WhatsApp in 2023. Our Mentor works in French text and Bamanankan voice. In our AI build log, we share how we collected and validated local-language data and built our WhatsApp support agent.

Native speakers translated the phrases, supervisors validated them, and we used a shared protocol to record them.

What the speech tests showed

We fine-tuned Whisper medium for Bamanankan using audio we curated and existing speech datasets. In August 2026, it reached 24% word error on our held-out data and 33% on an external benchmark.

How to read this evidence

Our 2025 model reached 64% on a different test set. We share the voice pipeline and test results in our playbook.

Read the sources

Read the primary source