Highdigenous Tech

language technology

Building Bamanankan speech data: why Kabakoo turned to linguistic expertise

By · · EN

Kabakoo shifted from open participation to dedicated Bamanankan linguists after encountering inconsistent translations and recording problems. Its dataset workflow combines real app conversations, expert validation, and French-language context during recording.

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

From open participation to Bamanankan linguists

Kabakoo shifted from an open community call to a dedicated group of Bamanankan linguists when translation and recording quality required specialist expertise. The initial selection form assessed participants’ proficiency in reading and writing Bamanankan, their spoken dialect, and their gender, with the aim of balanced representation.

Despite that structured selection, translations were inconsistent, recordings often contained background noise or unnatural intonations, and work progressed more slowly than expected. Kabakoo turned to linguists to ensure accuracy and consistency across translations and recordings. Its approach combines native-speaker translation with rigorous expert validation.

Building sentences from real app conversations

The phrases began with real interactions with Kabakoo’s AI system within the app. Kabakoo analyzed conversations, extracted key topics, and used them to generate natural, contextually rich sentences for the dataset.

On its custom translation platform, translators see a French sentence and enter its Bamanankan translation. They can skip difficult sentences. The project’s goal was to deliver 10 hours of high-quality recorded speech from native speakers.

Checking batches and limiting daily translation

After noticing that accuracy declined when translators worked on too many batches in a single session, Kabakoo capped translations at 100 sentences per translator per day.

A supervisor randomly checks 25% of the sentences in each submitted batch. If even one sentence is incorrect, the entire batch returns for correction with comments. That cycle repeats until quality standards are met. Random selection addresses variation across the working day: a translator may begin with high accuracy but produce lower-quality translations toward the end.

Keeping meaning visible during recording

In Bamanankan, ja can mean “to dry”, “to petrify”, or “the shade of a tree”, depending on pronunciation and context. Intonation is crucial to meaning.

Narrators therefore see the translated Bamanankan sentence alongside the original French text. Without that context, they can misinterpret meaning and produce unnatural phrasing. Adding the French sentence helped narrators adjust intonation and deliver the intended meaning accurately. Kabakoo describes this translation, validation, and recording process in Before Training, the Data!

The Kabakoo newsletter

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