Maloba

Speech corpus · Reference version

audios-lingala-annotatees-v3.2

18,767 segments, 58.13 h, 210 source recordings, harmonised transcriptions.

View on Hugging Face (opens in a new window)All corpora

How it was built

  • Recordings from the partner rural radio stations and from the Health Voice-a-thon, segmented and transcribed by annotators trained in Label Studio.
  • Rebuild (V2): audio re-extracted losslessly (FLAC 16 kHz mono), with an audio fingerprint and the identifier of the source recording.
  • Deduplication (V3) on the fingerprint; splits drawn by recording, so no audio is shared between train, validation and test.
  • Harmonisation (V3.1, V3.2) of the transcriptions under the Maloba convention; the original text is kept in its own column.

Figures

Splits of corpus V3.2 (source: Hugging Face dataset card)
SplitSegmentsSource recordings
Training15 175172
Validation1 79519
Test1 79719
Total18 767210

Total duration: 58,13 h. Segments from 0.3 to 25 s, FLAC 16 kHz mono.

The corpus, as charts

The three splits of the corpusThe splits are drawn by recording and not by segment: no audio is shared between training, validation and test. That is what makes the measurement honest.
04 0978 19512 29216 389Training15 175 segmentsValidation1 795 segmentsTest1 797 segments

Source : Fiche du corpus V3.2 · data

Show the chart data
SplitSegmentsDetail
Training15 175172 recordings · 47.93 h
Validation1 79519 recordings · 5.37 h
Test1 79719 recordings · 4.83 h
What harmonisation changedBefore harmonisation, 600 words were written in several ways across the corpora; afterwards, 290. No sentence is rewritten: only the majority spelling of ɔ and ɛ is applied, and the original text is kept.
0162324486648Before harmonisation600 motsAfter harmonisation290 mots

Source : Fiche du corpus V3.2

Show the chart data
StateWords written several ways
Before harmonisation600
After harmonisation290

Limits

  • The test set holds only 19 source recordings: compute confidence intervals by recording, not by segment.
  • Field speech, with noise and Lingala–French code-switching; Lingala spelling is not settled.
  • The number of distinct speakers is not established.

Protecting the speakers

The corpus holds human voices. It carries no name and no contact details. Seeking to identify speakers or to clone their voice is forbidden: that is a condition of access. To withdraw a recording, write to contact@congo-digital.com. Read the data governance.

Citation

Congo Digital Services (2026). MALOBA Lingala speech corpus v3.2 (audios-lingala-annotatees-v3.2). Hugging Face. NOODL-1.0 licence.