Version 3.2 of the audios-lingala-annotatees corpus is the project's reference version. It holds 18,767 audio segments and 58.13 hours of speech, drawn from 210 source recordings: broadcasts from the partner rural radio stations and the Health Voice-a-thon.
What changed from the earlier versions
- Rebuild (V2): each segment is re-extracted from its source recording and kept losslessly, with an audio fingerprint.
- Deduplication (V3): audio ingested several times under different identifiers is deduplicated. The train, validation and test splits are drawn on that fingerprint: no audio is shared between them.
- Harmonisation (V3.1 then V3.2): transcriptions receive a harmonised version under the Maloba writing convention. The original transcription is kept alongside.
Why leak-free splits
The earlier speech-recognition measurements were made on a test set that shared audio with the training set. Re-measuring on the V3.2 test set (1,797 segments, 19 recordings never seen) gives a word error rate of 44.17 % for fine-tuned Whisper. That is the figure we publish from now on.
Access
The corpus holds human voices. It is published under the NOODL-1.0 licence with controlled access: every request is reviewed, and the user undertakes not to seek to identify the speakers or to clone their voice.
By The Maloba team, September 29, 2026.