MALOBA Lingala Speech Corpus (ASR), v3.2

TECHNICAL DATASHEET — MALOBA Lingala Speech Corpus (ASR), v3.2

TYPE OF DATASET Speech dataset for automatic speech recognition (ASR): audio segments paired with Lingala transcriptions. This is NOT a translation dataset. French appears only as code-switching inside Lingala speech.

SOURCE AND COLLECTION Radio programmes in Lingala provided by Radio Rurale, the sole owner of the recordings, which consented to their use in the project and to the distribution of this dataset. 210 source recordings were segmented and transcribed by annotators following a written guideline (Label Studio): lower case, no punctuation, no tone marks, numbers written in words, tags for code-switching and non-speech noise.

SIZE AND FORMAT 18,767 segments, 58.1 hours. Audio: FLAC, 16 kHz, mono. Segment duration 0.3–25 s (median 9.1 s).

• train: 15,175 segments, 47.93 h, 172 recordings
• validation: 1,795 segments, 5.37 h, 19 recordings
• test: 1,797 segments, 4.83 h, 19 recordings

DEDUPLICATION AND SPLITS Each segment carries a fingerprint (empreinte_pcm, SHA-256 of the decoded samples). Splits are drawn on this fingerprint, so identical audio never appears in two splits. 1,959 segments (10.4 %) are repeated audio copies; they are kept but always grouped in the same split. The 58.1 h figure includes them.

TRANSCRIPTION HARMONISATION (deterministic; no human editing and no generative model)

1. Unicode look-alike characters fixed (e.g. Greek epsilon → ɛ, ͻ → ɔ), NFC normalisation.
2. Tone marks removed.
3. Annotation guideline rules applied.
4. ɔ/ɛ spelling set by majority vote from a frozen lexicon of 11,382 entries, built on train and validation only (no test leakage). The original transcription is kept in the column transcription_v3. The harmonised text is in transcription_harmonisee (also exposed as text). The column modifications / nb_modifications lists what was changed on each line.

COLUMNS audio, text, transcription_v3, transcription_harmonisee, modifications, nb_modifications, empreinte_pcm, id_enregistrement, duree_s, vu_train_v2, vu_validation_v2, dans_test_v2, eval_propre.

RECOMMENDED EVALUATION Use split == "test" and eval_propre == True. Apply the same harmonisation to model outputs before scoring. Report WER and CER, both raw and normalised. The test set has only 19 recordings: use bootstrap over recordings for confidence intervals. Spelling variation alone accounts for roughly 5 points of WER on the test references, so raw and normalised scores should not be mixed.

KNOWN LIMITATIONS

• Radio speech only; studio speech is over-represented. Field and telephone speech are not covered.
• The number of distinct speakers is not established; do not assume speaker-independent splits.
• Lingala and French are mixed in places.
• Accents were removed from some French words (e.g. émission → emission).
• A few non-speech tags remain (e.g. cough, music, laughter).
• The spelling convention has not been validated word by word by linguists.

VERSION HISTORY v2: rebuilt corpus. v3: deduplicated, splits redrawn on the fingerprint. v3.1: first harmonisation. v3.2: current reference (fixed ͻ → ɔ, 115 occurrences). Reference revision on Hugging Face: Congo-digital-service/audios-lingala-annotatees-v3.2 @ fed93c0e49fe2b2b0668dc4f085bb6c9546bff76

LICENCE AND CONTACT NOODL-1.0 (gated access). Contact: [email protected]

ACCESS The dataset is open to anyone who accepts NOODL-1.0, not only to CDS members. Access requests are approved automatically: this records who downloads the data and ensures every user accepts the licence, including attribution to Radio Rurale (the rights holder) and to the MALOBA project. Users outside Africa and developing countries must contact Congo Digital Services before any use. Access can be revoked if the licence is not respected.