Congolese Speech Radio Corpus (CSRC)

The Congolese Speech Radio Corpus (CSRC) is an unlabelled radio-speech corpus covering Lingala, Kikongo, and Tshiluba.

This Hugging Face release was prepared by Bantu Languages Initiative from the CSRC component of Speech Recognition Datasets for Congolese Languages. Its purpose is to make the radio archives easier to use for self-supervised speech learning, ASR pretraining, acoustic adaptation, language identification, and robust speech research on under-resourced Bantu languages.

This is not the original publication repository.

Repository structure

The repository contains a single train split. Each row corresponds to one 16 kHz mono audio segment and includes:

• the audio segment;
• its language;
• the original source recording;
• its absolute start and end positions in the source recording;
• parsed broadcast metadata when available;
• automatic signal-quality measurements;
• a hierarchical quality score.

The audio is stored only once. Researchers can select a stricter or broader subset by filtering the quality_score field.

Quality levels

Each accepted segment receives the highest quality level that it satisfies:

Quality level	Score	Intended use
challenging	1	Robust speech systems and training with realistic radio-domain difficulties
standard	2	General self-supervised learning, ASR pretraining, and domain adaptation
high_quality	3	Stricter automatically selected subset for more controlled experiments

The levels are cumulative:

• quality_score >= 1 returns every accepted segment;
• quality_score >= 2 returns standard and high_quality;
• quality_score >= 3 returns only high_quality.

high_quality is an automatically selected subset. It must not be interpreted as a manually verified gold corpus.

Radio-aware VAD and chunking

The source material consists of radio programmes. In this domain, presenters may speak continuously for long periods, with very short pauses between sentences. Background music, production beds, and broadcast compression can also make short pauses less visible to a voice activity detector.

For this reason, a neural VAD may return a single speech region lasting well over one minute, even when the recording contains several natural sentence or topic boundaries. The maximum segment duration is therefore not enforced inside the VAD itself.

The release uses the following two-stage strategy:

1. Voice activity detection identifies speech-dominant regions in the original radio recordings.
2. Deterministic post-VAD chunking subdivides long VAD regions without running the VAD again.

Long regions are split into balanced subsegments so that no published segment exceeds the most permissive duration ceiling of 50 seconds. The absolute timeline is preserved: segment_start_s and segment_end_s always refer to positions in the original source recording.

This strategy avoids discarding many hours of valid radio speech simply because a presenter spoke for a long time without a sufficiently long silence.

Intro and outro exclusion

To reduce station identifiers, opening themes, credits, and programme endings, the following source zones are excluded from every quality level:

• the first 90 seconds of each source recording;
• the final 30 seconds of each source recording.

These rules are based on inspection of the source archive and are applied before final quality assignment.

Quality policy

Quality levels use nested duration and signal-integrity thresholds.

Level	Duration	Minimum RMS	Minimum peak	Maximum clipping ratio	Maximum near-zero ratio
challenging	4–50 s	-55 dBFS	0.003	2.0%	60%
standard	5–35 s	-45 dBFS	0.008	1.0%	40%
high_quality	6–30 s	-38 dBFS	0.015	0.3%	25%

These checks are deliberately hierarchical:

• challenging maximizes usable quantity while excluding clearly unusable audio;
• standard provides a balanced quality/quantity trade-off;
• high_quality applies the strictest automatic filtering.

The signal-level checks do not replace a dedicated speech/music classifier. In particular, some challenging examples may still contain background music, production beds, overlapping speech, or archive artefacts.

Loading the dataset

from datasets import load_dataset

ds = load_dataset(
    "BantuLanguagesInitiative/CSRC",
    split="train",
)

Select the complete accepted release:

challenging = ds.filter(
    lambda example: example["quality_score"] >= 1
)

Select the standard cumulative subset:

standard = ds.filter(
    lambda example: example["quality_score"] >= 2
)

Select the strictest subset:

high_quality = ds.filter(
    lambda example: example["quality_score"] >= 3
)

Filter by language:

lingala = ds.filter(
    lambda example: example["language"] == "ln"
)

Dataset statistics

Total accepted segments: 97,779

Total accepted duration: 690.35 hours

minimum_quality	language	segments	hours	mean_duration_s
challenging	Kikongo	32703	233.86	25.74
challenging	Lingala	33402	232.97	25.11
challenging	Tshiluba	31674	223.52	25.41
standard	Kikongo	32659	233.78	25.77
standard	Lingala	33366	232.91	25.13
standard	Tshiluba	31639	223.47	25.43
high_quality	Kikongo	32595	233.47	25.79
high_quality	Lingala	33304	232.7	25.15
high_quality	Tshiluba	31584	223.2	25.44

Data fields

• segment_id: unique segment identifier.
• audio: 16 kHz mono audio segment.
• language: language code.
• language_name: language name.
• source_id: deterministic identifier of the original radio recording.
• source_filename: original archive filename.
• broadcast_date: date parsed from the source filename when available.
• date_parse_rule: rule used to parse the date.
• time_slot: inferred broadcast time slot when available.
• program_type: inferred programme type when available.
• segment_index: index of the segment within the source recording.
• segment_start_s: absolute segment start time in the source recording.
• segment_end_s: absolute segment end time in the source recording.
• source_duration_s: duration of the source recording.
• duration_s: duration of the segment.
• sampling_rate: audio sampling rate.
• num_channels: number of audio channels.
• rms_dbfs: RMS level in dBFS.
• peak: maximum absolute sample value.
• clipping_ratio: proportion of near-clipped samples.
• zero_ratio: proportion of near-zero samples.
• quality_score: cumulative quality score from 1 to 3.
• quality_level: highest quality level reached.
• challenging_flags: reasons the segment does not satisfy the challenging policy.
• standard_flags: reasons the segment does not satisfy the standard policy.
• high_quality_flags: reasons the segment does not satisfy the high_quality policy.

The repository also includes full processing manifests under metadata/, including rejected segments and their rejection reasons.

Intended uses

CSRC is intended for:

• self-supervised speech representation learning;
• ASR pretraining;
• acoustic and domain adaptation;
• robust speech recognition;
• language identification;
• speech segmentation research;
• research on under-resourced Congolese and Bantu languages.

The corpus is unlabelled and is not a supervised ASR benchmark without additional transcription.

Relation to LRSC

The Lingala Read Speech Corpus (LRSC) is the labelled read-speech component of the same original publication.

LRSC and CSRC are complementary:

• LRSC provides a smaller supervised corpus with paired audio and transcriptions;
• CSRC provides larger-scale, naturally occurring radio speech suitable for self-supervised learning and radio-domain adaptation.

Source and attribution

The original dataset was published as:

Kimanuka, Ussen; wa Maina, Ciira; Büyük, Osman (2023).
“Speech Recognition Datasets for Congolese Languages.”
Mendeley Data, Version 1.
DOI: 10.17632/28x8tc9n9k.1

This Hugging Face release is a repackaging and preprocessing effort by Bantu Languages Initiative to improve accessibility for the African speech and NLP research community.

License

The original dataset is distributed under Creative Commons Attribution 4.0 International (CC BY 4.0).

This repackaged release is also distributed under CC BY 4.0. Users must provide appropriate attribution to the original authors and dataset.

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}
}

Suggested acknowledgement:

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

Limitations

• Segmentation and quality classification are automatic.
• Some speech/music overlap may remain, especially in the challenging subset.
• The quality levels are based on duration and signal-level measurements, not manual listening.
• Broadcast dates and programme metadata are parsed from filenames and may be uncertain.
• The corpus does not include speaker diarization or speaker identity.
• The corpus does not include transcriptions.
• Radio archives may contain compression artefacts, overlapping speakers, jingles, music beds, and varying acoustic conditions.