Datasets:

BantuLanguagesInitiative
/

LRSC

license: cc-by-4.0
language:
  - ln
pretty_name: Lingala Read Speech Corpus (LRSC)
task_categories:
  - automatic-speech-recognition
tags:
  - lingala
  - bantu-languages
  - congolese-languages
  - low-resource
  - automatic-speech-recognition
  - read-speech
  - speech
size_categories:
  - 1K<n<10K

Lingala Read Speech Corpus (LRSC)

This repository provides a Hugging Face-ready repackaging of the Lingala Read Speech Corpus (LRSC), the labelled Lingala subset of Speech Recognition Datasets for Congolese Languages.

The goal of this repackaging by Bantu Languages Initiative is to make the annotated LRSC subset easier to load, inspect, cite, and use with the Hugging Face datasets library for ASR research on Lingala and under-resourced Bantu languages.

This is not the original publication repository. The original dataset was published on Mendeley Data by Kimanuka, wa Maina and Büyük.

Dataset summary

Language: Lingala (ln)
Task: Automatic Speech Recognition (ASR)
Speech type: read speech
Total duration in this HF package: 5.01 hours
Number of utterances: 2940
Number of speakers detected from filenames: 31
Audio sampling rates in original files: 16000 Hz: 2940
Channels in original files: 1 channel(s): 2940

Splits

The original LRSC release contained train/validation style manifests. For easier use in modern Hugging Face ASR workflows, this repository provides deterministic train, validation, and test splits.

When possible, the split is speaker-level, so the same detected speaker ID does not appear in multiple splits.

split	utterances	speakers	hours	ratio_hours
test	463	3	0.863	0.172428
train	1924	25	3.221	0.643556
validation	553	3	0.921	0.184016

Speaker metadata

The speaker_id is extracted from the audio filename prefix. The gender field is manually annotated from representative audio samples and should be interpreted as perceived voice category, not as a biological attribute.

gender	speakers	utterances	hours
female	20	1892	3.255
male	11	1048	1.75

Character inventory

The original dict.ltr.txt file is included in this repository as dict.ltr.txt, because the corpus uses characters and diacritics beyond a minimal Latin alphabet. This is important for CTC-style ASR experiments and for preserving Lingala tonal/orthographic information.

Characters found in dict.ltr.txt:

’, x, a, w, à, f, -, q, 5, ç, d, î, j, e, 0, g, s, o, c, ', h, 3, t, l, ǎ, r, ε, ê, 8, y, n, |, u, ɔ, z, ɛ, k, m, å, v, i, p, b

Data fields

Each example contains:

utterance_id: unique utterance identifier, derived from the audio filename.
audio: audio file decoded by Hugging Face datasets.
transcription: original text transcription.
speaker_id: speaker identifier extracted from filename.
gender: perceived voice category (female, male, unknown).
language: ISO-like language code, here ln.
language_name: Lingala.
duration_s: duration in seconds.
sampling_rate: original file sampling rate.
num_channels: number of channels in the original file.
original_split: original LRSC split before this HF repackaging.

Relation to other Lingala ASR resources

LRSC is complementary to larger African speech datasets such as google/WaxalNLP. While WaxalNLP provides a much larger multilingual ASR/TTS resource, LRSC is useful as a compact supervised Lingala benchmark with carefully paired audio and text. It can be especially useful for quick experiments, sanity checks, low-resource ASR baselines, and evaluation of models trained on larger Lingala speech resources.

License

The original dataset page indicates a CC BY 4.0 license. This repackaging is therefore distributed under CC BY 4.0.

Users must cite the original authors and dataset when using this corpus.

Citation

Please cite the original dataset:

@dataset{kimanuka_2023_congolese_speech,
  author    = {Kimanuka, Ussen and wa Maina, Ciira and Büyük, Osman},
  title     = {Speech Recognition Datasets for Congolese Languages},
  year      = {2023},
  publisher = {Mendeley Data},
  version   = {V1},
  doi       = {10.17632/28x8tc9n9k.1}
}

Limitations

LRSC is a small read-speech corpus and should not be treated as fully representative of all Lingala accents, dialects, spontaneous speech, or noisy real-world speech.
Speaker IDs are inferred from filenames.
The gender field is manually added in this repackaging as perceived voice category and may be uncertain.
The train/validation/test splits are newly generated for Hugging Face convenience; the original_split column preserves the original split provenance.

This Hugging Face version was prepared by Bantu Languages Initiative to improve accessibility for the ASR and African NLP research community.