Congo-digital-service/audios-lingala-annotatees-v2

Annotated Lingala Audio — v2

Superseded. This version is kept unchanged as the training data of the fine-tuned Whisper model. For new work, use audios-lingala-annotatees-v3.2: same recordings, deduplicated on the audio fingerprint, with harmonised transcriptions.

Annotated Lingala speech for open automatic speech recognition research and for fine-tuning speech models.

This release is a full reconstruction of the corpus from its source recordings and annotations. It supersedes Congo-digital-service/audios-lingala-annotatees, which is deprecated — see Relationship to the previous release below.

Datasheet

Datasheet following the template shared by UNDP (after Gebru et al., Datasheets for Datasets, and Bender & Friedman, Data Statements for NLP).

Item	Description
Core purpose	Train and evaluate automatic speech recognition (ASR) for Lingala. Training data of the fine-tuned Whisper model.
Motivation	Lingala has very little open speech data. Built by CDS in the MALOBA project, with UNDP Republic of Congo and the Ministry of Posts, Telecommunications and the Digital Economy.
Curation rationale	Full reconstruction of the corpus from the source recordings and annotations, with the audio embedded, so that every segment can be traced and the corpus split by source recording.
Language / variety	Lingala (ln) as broadcast by Radio Rurale of the Republic of Congo, with French borrowings. Transcriptions as written by the annotators, not harmonised.
Speaker demographics	Presenters, guests and callers heard on air. The number of distinct speakers and their age, sex and region were not recorded.
Annotator demographics	A team of Lingala-speaking annotators, mostly graduates of the Faculté des Lettres et des Langues Vivantes, Université Marien Ngouabi (UMNG), trained in AI and IT and paid for their work by CDS. Identities are not published.
Speech situation	Radio broadcasts: spontaneous and semi-prepared speech intended for a general audience.
Composition	18,767 segments (58 h) from 210 source recordings; 1,959 segments flagged as audio duplicates (kept). See Statistics.
Collection process	Recordings provided by Radio Rurale; segmented by CDS; transcribed in Label Studio; corpus rebuilt from the annotation bucket. See How the corpus was built.
Preprocessing / cleaning	Audio re-extracted from the source recordings at 16 kHz mono, with a PCM fingerprint per segment. Transcriptions kept as written.
Uses	ASR training and evaluation; reproduction of the Whisper model. For new work, use v3.2. Not allowed: identifying speakers, voice cloning, speaker identification.
Distribution	Hugging Face, gated access. Licence NOODL-1.0, full text in LICENSE.md.
Maintenance	Kept unchanged as a provenance record. Maintained by CDS (contact@congo-digital.com); corrections are published in later versions.
Legal & ethical compliance	Radio Rurale, sole owner of the recordings, consented to their use in the MALOBA project and to their distribution. No speaker names or contact details are included.
Provenance	Radio Rurale broadcasts → Label Studio annotations (MALOBA) → reconstruction (this version) → deduplication (v3) → harmonisation (v3.1, v3.2).

What this dataset contains

Each row is one annotated speech segment, carrying the audio itself, its transcription, and the metadata needed to trace it back and to split the corpus correctly.

Column	Description
audio	The waveform, embedded in the dataset. No external files, no paths to resolve.
text	Human transcription of the segment
language	Declared language — ln for every row
id_segment	Identifier of this segment
id_enregistrement	Source recording the segment was cut from — the key to split on
id_tache, id_annotation, id_region	Identifiers in the annotation pipeline, kept for traceability
id_source	Historical clip name of the previous release
source_audio	Provenance key of the source recording
debut_s, fin_s, duree_s	Position of the segment inside its source recording, and its duration
etiquette	Annotation label attached to the region
empreinte_pcm	Fingerprint of the decoded samples
doublon_audio	Flag: this segment shares its fingerprint with another. Flagged, never removed.

Timing fields and identifiers are published on purpose. The previous release kept only audio, text and language; without the source recording identifier, the corpus could not be split correctly, and without fingerprints, repeated transcriptions could not be told apart from genuine duplicates.

Mel spectrograms and tokenised labels are deliberately not published. They are model-specific pre-computations, they made the previous release four times heavier than the audio it was derived from, and they made the corpus unusable by any architecture other than the one they were computed for. Compute them at training time instead.

Statistics

Metric	Value
Segments	18,767
Distinct source recordings	210
Segments sharing a source recording	18,767 (100%)
Segments flagged as audio duplicates	1,959
Total duration	58 h 07 min 56 s
Average segment duration	11.15 seconds
Sampling rate	16 kHz, mono

Splits

Split	Segments	Duration	Purpose
train	15,001	47 h 13 min 35 s	Model fitting. The only split that may be augmented.
validation	1,884	5 h 42 min 46 s	Model selection, early stopping, hyper-parameter choice.
test	1,882	5 h 11 min 36 s	Held out. Measured once, on the final model.

Splits are grouped by source recording (id_enregistrement), not by segment. A large share of segments comes from a recording that produced at least one sibling; splitting at segment level would place the same speaker, recorded in the same acoustic conditions, on both sides of the boundary, and would make any reported error rate optimistic.

This three-way split exists so that model selection never touches the test set. Reporting a figure obtained by early-stopping on the same split used to publish it is not a measurement.

Loading

from datasets import load_dataset

corpus = load_dataset("Congo-digital-service/audios-lingala-annotatees-v2")
example = corpus["train"][0]

example["audio"]["array"]          # the waveform
example["audio"]["sampling_rate"]  # 16000
example["text"]                    # the transcription

How the corpus was built

The recordings are programmes broadcast by Radio Rurale of the Republic of Congo, the sole owner of the recordings and of the rights on them. They were segmented by CDS and transcribed by human annotators through a Label Studio pipeline. Each published row carries the identifier of its source recording and of the annotation region it came from, so that any transcription can be traced back to the original audio.

Radio Rurale of the Republic of Congo is the only source of the recordings.

Content and context

• Language: Lingala (ln)
• Topics: daily life, family, relationships, work and trade, religion, travel, health, education
• Register: spoken, everyday, conversational — urban Lingala, with borrowings from French

Known limitations

These are stated here rather than left to be discovered.

Unicode is not harmonised. The corpus mixes the Greek epsilon ε (U+03B5) with the Lingala ɛ (U+025B). The two are visually identical and informatically distinct, and both occur in the transcriptions. The corpus deliberately keeps its original spelling; normalisation belongs to the measurement step, not to the data. Any word error rate computed on this corpus should fold these characters together before scoring, and report both figures.

Spelling variants are unresolved. Forms such as po na and mpo na, or nyɔnso and nyonso, coexist. These are genuine orthographic variants of a language whose written norm is still settling, not encoding noise.

Repeated transcriptions and duplicates. Identical transcriptions are common in this corpus — some are genuine repetitions of everyday phrases, some are the same audio segment appearing twice. The two are told apart by a fingerprint of the decoded samples: same transcription and same fingerprint is a duplicate, same transcription with different fingerprints is not. Duplicates carry the doublon_audio flag and are kept in the corpus, so that the decision to exclude them stays with whoever uses it.

Speaker counts are not established. The number of distinct speakers is under verification by diarisation and has not been published here, rather than estimated.

Quality is under independent review. A sample of this corpus is being evaluated by external annotators who did not take part in the original annotation, to produce a measured validity rate. Results will be published alongside this card.

Relationship to the previous release

audios-lingala-annotatees is deprecated and should not be used for new work. It published a column of relative audio paths without the corresponding bytes: the audio it referred to was never included, and its 27.5 GB consisted entirely of pre-computed spectrograms. It is kept online, unchanged, solely as the provenance record of the models that were trained from it.

Metrics published against the previous release are not comparable with metrics computed on this one.

Creators

Congo Digital Services (CDS SARL) — https://www.congo-digital.com/

In collaboration with:

• the MALOBA community — https://maloba.congo-digital.com/
• Radio Rurale
• the Faculté des Lettres et des Langues Vivantes, Université Marien Ngouabi (UMNG)
• UNDP Republic of Congo
• the Ministry of Posts, Telecommunications and Digital Economy of the Republic of Congo

License

This dataset is published under the Nwulite Obodo Open Data License (NOODL-1.0), a license designed for equitable sharing of African datasets.

This dataset, audios-lingala-annotatees-v2 (Congo-digital-service/audios-lingala-annotatees-v2), was created by Congo Digital Services (CDS SARL) (https://www.congo-digital.com/) in collaboration with the MALOBA community (https://maloba.congo-digital.com/). It is licensed under the Nwulite Obodo Open Data License (https://licensingafricandatasets.com/nwulite-obodo-license).

In exchange for use of this dataset, users from developing countries are required to provide: Use of the dataset (no additional benefit required). Users from high-income countries or commercial entities are additionally required to publicly acknowledge and credit the Maloba Project (UNDP Republic of Congo — language digitalisation initiative) in any publication, product, model, or output derived from this dataset. To fulfil this requirement, contact contact@congo-digital.com or visit https://www.congo-digital.com/contact.

Citation

@misc{cds2026lingala,
  title  = {Annotated Lingala Audio — canonical corpus},
  author = {{Congo Digital Services (CDS SARL)} and {MALOBA Project}},
  year   = {2026},
  howpublished = {\url{https://huggingface.co/datasets/Congo-digital-service/audios-lingala-annotatees-v2}},
  note   = {In partnership with UNDP Republic of Congo and the Ministry of Posts, Telecommunications and Digital Economy}
}