appropriate technology
Building a Bambara AI mentor around how learners speak and read
Kabakoo pairs French text with Bambara audio to reflect learners’ language preferences. Its first-year experience offers practical guidance on bilingual design, speech-model evaluation and tracking use beyond launch.
Start with learners’ spoken and written languages
Kabakoo’s Bambara AI Mentor uses French text with Bambara audio: learners prefer Bambara for oral communication but often find French easier for reading and comprehension. For a learning system, the practical starting point is to reflect that combination rather than choose one language for every interaction.
The mentor was designed and piloted entirely in Mali by the Kabakoo team. In its account of the pilot’s first year, Kabakoo describes Bambara, also known as Bamanankan, as a lingua franca for millions in Mali, Burkina Faso, Côte d’Ivoire, Guinea and Senegal.
Pair text and audio around everyday language use
Design for both languages together. Kabakoo’s learners naturally mix Bambara and French in daily life, reinforcing the need for a system that supports both. The bilingual approach pairs the language learners often find easier to read with the language they prefer to speak.
That pairing is the concrete design choice to carry forward from this experience: French text with Bambara audio, rather than a single-language interface. Explore Kabakoo’s AI learning system for more about its approach.
Measure speech performance and identify the data gap
Track model performance alongside the audio available to improve it. Using OpenAI’s Whisper architecture, Kabakoo reduced its Bambara speech model’s Word Error Rate from 0.98 to 0.64. The reported improvement sits alongside a specific constraint: only 10 hours of high-quality Bambara audio data.
Kabakoo identifies a need for an additional 90 hours to build a truly useful and scalable model. Keep those measures together when assessing progress: the reported Word Error Rate, the available high-quality audio and the additional data still needed.
Follow continued use, not only the launch peak
Review engagement beyond the initial activity. During Bamako.AI #2 in May and June 2024 following launch, interaction with the mentor peaked at 1,452 conversations, followed by 1,122 conversations one month later.
Engagement then decreased significantly through 2025, while a smaller, consistent user base remained. Preserve both observations when reviewing the pilot: continued activity and a substantial decline from the peak. Alongside bilingual design and speech-model measurement, that usage trajectory belongs in the assessment of the mentor’s first year.