You need to agree to share your Hugging Face information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
CamSpoken is a research dataset from LAB 237, an independent AI research lab in Cameroon.
By accessing this repository, you agree to the following conditions:
- Research Use Only: This resource is restricted to non-commercial research use (CC-BY-NC-4.0).
- Citation: You agree to cite the relevant technical report in any derivative work or publication.
- Attribution: Proper acknowledgment must be given to LAB 237.
Log in or Sign Up to review the conditions and access this dataset content.
Dataset Card for CamSpoken
Dataset Summary
CamSpoken is an umbrella repository for spoken-language resources in oral-first Cameroonian language varieties, released by LAB 237, an independent AI research lab in Cameroon. Each variety and task is a separate configuration (config), so new varieties and new tasks can be added without changing what is already released.
The current release (v1.0) contains CamSpokenNLI, a pilot spoken Natural Language Inference (NLI) resource for Cameroon Pidgin English (Kamtok/Wes Cos, ISO 639-3: wes) and Camfranglais (a French-English-Pidgin urban youth vernacular; Glottolog: camf1234, no ISO 639-3 code assigned). To our knowledge, this is the first spoken NLI resource for Cameroon Pidgin, and the first NLI resource of any kind, spoken or text, for Camfranglais.
Each NLI config contains 114 samples, each a (premise audio, premise text, hypothesis text, label) tuple, with an English gloss of the premise and hypothesis so the dataset can be checked by readers who do not speak either variety.
This is explicitly a pilot release (v1.0), not a benchmark. At this scale it is too small to support strong statistical claims. It is released to catalyze larger-scale collection, by this lab or by others. See Limitations below.
Available Configs
Configs are named {variety}_{task}.
| Config name | Variety | Task | Samples (train) |
|---|---|---|---|
pidgin_nli |
Cameroon Pidgin English (Kamtok) | Spoken NLI | 114 |
camfranglais_nli |
Camfranglais | Spoken NLI | 114 |
How to Use
Specify config_name when loading.
This dataset is gated. Accept the terms on this page, then log in with hf auth login (or huggingface_hub.login()) before loading.
from datasets import load_dataset
# Cameroon Pidgin (Kamtok), NLI
pidgin = load_dataset("lab-237/camspoken", "pidgin_nli", split="train")
# Camfranglais, NLI
camfranglais = load_dataset("lab-237/camspoken", "camfranglais_nli", split="train")
# Reproduce the paper exactly by pinning the release tag
pidgin_v1 = load_dataset("lab-237/camspoken", "pidgin_nli", split="train", revision="v1.0")
example = pidgin[0]
print(example["premise_text"], "->", example["hypothesis_text"], "|", example["label"])
# Decode audio (requires: pip install datasets[audio] torchcodec)
samples = example["premise_audio"].get_all_samples()
print(samples.data.shape, samples.sample_rate)
To combine both varieties for pooled training or evaluation:
from datasets import concatenate_datasets
combined = concatenate_datasets([pidgin, camfranglais])
About LAB 237
CamSpoken is a release from LAB 237, an independent, research-driven initiative advancing artificial intelligence, natural language processing, speech technology, data science, and open research resources relevant to Cameroon and similar multilingual settings. It is not an organization built around a single dataset. Current and planned areas of work include:
- Spoken and text NLI for Cameroon Pidgin and Camfranglais (CamSpokenNLI, this release)
- Speech recognition and other speech resources for oral-first Cameroonian varieties
- Text-to-SQL and semantic parsing for low-resource settings
- Code-mixed NLP more broadly
- Small language models (SLMs) suited to low-resource deployment
- Other AI datasets, research benchmarks, and practical tools addressing research challenges in Cameroon, across Africa, and beyond
Expect further datasets, models, and tools from LAB 237; this release is one project among several.
Data Organization
CamSpoken v1.0 contains two NLI configs and a total of 228 pilot spoken NLI examples (114 per variety).
| Config | Variety | Task | Examples | Contents |
|---|---|---|---|---|
pidgin_nli |
Cameroon Pidgin English | Spoken NLI | 114 | Premise audio, premise text, hypothesis text, English glosses, label |
camfranglais_nli |
Camfranglais | Spoken NLI | 114 | Premise audio, premise text, hypothesis text, English glosses, label |
Configs are named {variety}_{task}. New varieties and tasks are added as new configs (see Roadmap).
Dataset Structure
Repository layout on the Hub:
lab-237/camspoken
βββ README.md
βββ pidgin_nli/
β βββ train-*.parquet (audio embedded)
βββ camfranglais_nli/
β βββ train-*.parquet (audio embedded)
βββ source/ (original files, for exact reproduction)
βββ pidgin.csv
βββ camfranglais.csv
βββ audio/
βββ pidgin/
β βββ wes_00001.wav
β βββ ... (wes_00114.wav)
βββ camfranglais/
βββ camf_00001.wav
βββ ... (camf_00114.wav)
Use load_dataset with the config names above for normal use. The source/ folder holds the original CSV and WAV files used for the paper's experiments (see Reproducing the Paper).
Data Fields
| Field | Type | Description |
|---|---|---|
id |
string | Unique, permanent sample ID, e.g. wes_00001, camf_00001 |
premise_text |
string | The premise, in Cameroon Pidgin or Camfranglais |
premise_text_en |
string | English gloss of the premise |
premise_audio |
audio | 16 kHz mono WAV recording of the premise |
hypothesis_text |
string | The hypothesis, in the same variety as the premise |
hypothesis_text_en |
string | English gloss of the hypothesis |
label |
string | One of entailment, contradiction, neutral |
The labels are:
entailment: the hypothesis must be true if the premise is true.contradiction: the premise and hypothesis cannot both be true.neutral: the premise does not establish either entailment or contradiction.
The English glosses are approximate references for understanding and research. They are not independent English NLI examples.
Items with the same number in the two configs are scenario-matched (for example, wes_00007 and camf_00007)
Data Splits
v1.0 ships a single train split per config (114 examples each). There is no official train/test partition in v1.0; train is only the default Hub split name. The paper uses a shared stratified 5-fold item-level split, and the fold assignments are in results/shared_item_folds.csv in the code repository. Dedicated dev/test splits are planned for v2.0 alongside the larger sample size. See Roadmap.
Reproducing the Paper
The experiments in the CamSpokenNLI paper (XLM-RoBERTa-base fine-tuning and zero-shot Whisper) were run on the original files in the source/ folder, not on the parquet configs. To reproduce them, download pidgin.csv, camfranglais.csv and audio/ from source/ and follow the instructions in the code repository (release v1.0). Use the files exactly as downloaded: the scripts check both CSVs against SHA-256 hashes.
For ordinary use (training, evaluation, browsing), load the configs with load_dataset as shown above.
Audio Recordings
Only the premise sentences are recorded. Hypotheses are provided as text and have no audio in this version.
- Format: WAV, 16-bit PCM
- Sample rate: 16 kHz
- Channels: Mono
- Duration: 2.0-6.8 seconds after trimming
- Content: One sentence per audio file
- Speaking style: Natural conversational pace and consistent volume
- Padding: 0.1 seconds of silence before speech and 1.0 second after speech
- Recording environment: Quiet room, with a phone microphone approximately 15 cm from the speaker's mouth
Dataset Creation
Source Data
Sentences reflect everyday life in Buea/Molyko, Cameroon: electricity, water, transport, market, school, mobile money, health, weather, church, farming, and rent. Written and recorded by the dataset author, a native speaker of both varieties.
Annotations
Labels were assigned by the author alone in v1.0; there is no independent inter-annotator agreement score yet. A second independent annotator and a reported Cohen's kappa are planned for v2.0.
Positionality
All v1.0 recordings are from a single speaker (the author, male, Buea). This is a limitation, since findings may not generalize across Cameroon Pidgin's other recognized varieties (Grafi, Liturgical, Francophone, Limbe, Bororo) or across other regional forms of Camfranglais. It is also a strength: labels reflect firsthand native-speaker judgment rather than translation or secondhand elicitation.
Intended Uses
CamSpoken v1.0 may be useful for:
- Exploratory natural language inference research in Cameroon Pidgin English and Camfranglais.
- Studying semantic inference in under-resourced language varieties.
- Investigating transfer between the two varieties.
- Comparing variety-specific text models with English-gloss reference conditions.
- Examining the limitations of existing multilingual language models on locally grounded examples.
- Exploratory ASR evaluation using the premise recordings and corresponding premise text.
The dataset is intended for research and experimentation. Its size and design do not support treating it as a definitive benchmark of NLI ability.
Limitations
- Pilot scale (114 per variety): too small for statistically strong claims; results should be read as illustrative, not conclusive.
- Single speaker, single accent.
- Self-annotated, with no inter-annotator agreement in v1.0.
- The two varieties are scenario-matched in v1.0, so models may exploit parallel alignments across them. This will be removed in v2.0.
- No reliable published population estimate exists for Camfranglais speakers; none is asserted here.
- Phone-recorded audio, not studio quality. It is adequate for NLI baselines but may need further processing for downstream TTS work.
Roadmap
Planned for v2.0 and beyond:
- Speakers, annotation, and splits: multiple speakers across demographic groups and regions, varied devices, independent annotators with reported agreement (Cohen's or Fleiss' kappa), and held-out dev/test partitions.
- Unmatched contexts: distinct scenarios for Camfranglais that decouple it from the Pidgin subset, with hypotheses written by several annotators and labeled blind, and negation balanced across labels.
- ASR configs: speech-recognition configs (for example,
pidgin_asr,camfranglais_asr) reusing existing recordings plus new ones, with unified transcription conventions. - End-to-end spoken NLI: spoken hypotheses alongside spoken premises, for inference directly from speech without intermediate transcription.
- Additional varieties: other oral-first Cameroonian varieties, each added as new configs named
{variety}_{task}.
Releases are identified by git tags (v1.0, v2.0, ...). Repository and config names do not carry version numbers.
Licensing Information
Released under cc-by-nc-4.0. Attribution appreciated via the citation below. The accompanying code and result files are released separately under the PolyForm Noncommercial License 1.0.0 (see the code repository).
Citation Information
If you use this resource, please cite the paper:
@misc{ebika2026camspokennli,
title = {{CamSpokenNLI}: A Pilot Resource for Cross-Variety Transfer Evaluation in {Cameroon Pidgin English} and {Camfranglais}},
author = {Ebika, Tiani Pekins},
year = {2026},
eprint = {2611.17573},
archivePrefix = {arXiv},
primaryClass = {cs.CL},
doi = {10.48550/arXiv.2611.17573},
url = {https://arxiv.org/abs/2611.17573}
}
Conflict of Interest Disclosure
This dataset is being built alongside LocalHands Africa, a voice-first service-exchange platform the author is developing. This connection is disclosed directly here and in the accompanying paper; readers should weigh the dataset's academic claims independently of its commercial motivation.
- Downloads last month
- 27