Our journey wasn't without obstacles. Here are some of the biggest lessons we learned:

Community Participation vs. Linguistic Expertise

Initially, we aimed for a community-driven approach. for both translation and vocalization. We launched an open call, inviting participants through a selection form that assessed their proficiency in reading and writing Bamanankan, their spoken dialect, and their gender to ensure balanced representation. However, despite this structured selection process, we encountered major quality issues as well as a lack of efficiency. Translations were inconsistent, and audio recordings often contained background noise or unnatural intonations. Another challenge was the slower-than-expected pace of work, which did not align with our production timeline.

These challenges made it clear that maintaining high-quality data required linguistic expertise rather than open participation. As a result, we shifted to working with a dedicated group of Bamanankan linguists to ensure the accuracy and consistency of both translations and recordings.

Fatigue Affects Quality

Initially, translators could complete as many sentences as they wanted per day. However, we observed a decline in accuracy after a certain threshold. Limiting batches to 100 sentences per day significantly improved consistency.

Context Matters for Natural Translations

Working with a low-resource language like Bamanankan, we found that intonation is crucial for meaning. The word ja, for example, can mean "to dry", "to petrify" or "the shade of a tree", depending on pronunciation and context. Without the original French text, narrators can often misinterpret meaning, leading to unnatural phrasing. Adding the French sentence during recording helped them adjust intonation and ensure accurate vocal delivery.

Validation Efficiency

The random validation method quickly highlights systematic errors, enabling prompt corrective action. We chose this method because a translator may start their day with high accuracy but produce lower-quality translations toward the end. A randomized selection ensures we capture a fair representation of the overall quality.

Interface Improvements

Translator feedback led to notable enhancements in our platform's design, making it more user-friendly and reducing errors.

For example, given the rigorous validation process, translators requested a feature to review their submitted work before sending it for supervisor validation. We implemented an interface where they can see all their translations in one place, edit them, and submit with confidence.

Recording Environment Impacts Sound Quality

Some early audio recordings had background noise or inconsistent volume. Providing a r helped standardize sound quality.