Highdigenous Tech

language technology

Bamanankan speech recognition: paid linguists, validated audio and model evaluation

By · · EN

Eight paid Bamanankan linguists helped Kabakoo build ten hours of validated audio. Its fine-tuned Whisper medium model reached 24% word error on held-out data and 33% on an external Bamanankan test set in August 2026.

A man wearing a colorful patterned shirt speaks to the camera in an indoor setting with a curtain background; a subtitle reads 'j'enseigne le Bamanankan et l'allemand au lycée' and the Kabakoo logo appears at the top.
Kabakoo Academies

Eight paid Bamanankan linguists helped Kabakoo build ten hours of validated audio. Kabakoo’s speech-recognition work combines specialist translation, narration and sampled review with model evaluation on held-out data and an external Bamanankan test set.

1. Learners in Bamako preferred oral explanations

At a vocational training center in Bamako, Kabakoo asked 40 learners how they wanted to learn. Most preferred oral explanations in Bamanankan. Ninety-two percent of those surveyed had difficulty reading Bamanankan, and half could not read it at all.

Kabakoo’s WhatsApp Mentor works in French text and has accepted Bamanankan voice messages since March 2024.

2. Linguists and narrators built the audio

The linguists translated on a platform Kabakoo built for the job. Narrators recorded with the original French in view to get the pronunciation and meaning right. The workflow brought together translation by Bamanankan specialists and narration with access to the French original during recording.

3. Supervisors sampled each batch for errors

Supervisors reviewed a random quarter of each batch of Bamanankan audio. When they found an error, they sent the batch back for correction. Review and correction were part of the workflow alongside translation and narration.

4. Kabakoo evaluated the model on separate test sets

Kabakoo fine-tuned Whisper medium with curated audio and existing Bamanankan speech datasets. In August 2026, the model reached 24% word error on held-out data and 33% on an external Bamanankan test set.

Kabakoo also works with the AI Accountability Lab at Trinity College Dublin on Bamanankan language AI. Together, they developed a research concept for testing how speech recognition works across speakers and using those findings to guide data collection with local linguists.

Read Kabakoo’s Bamanankan speech recognition account for the connected story of learner preferences, audio production and model testing.

The Kabakoo newsletter

Research, field notes, and stories from Kabakoo in your inbox.