README.md

---
language:
- en
- pcm
license: cc-by-4.0
task_categories:
- automatic-speech-recognition
tags:
- nigerian-pidgin
- speech
- africa
- low-resource
- naija
size_categories:
- 1K<n<10K
pretty_name: Pidgin ASR Combined
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: validation
    path: data/validation-*
  - split: test
    path: data/test-*
---

# Pidgin ASR Combined

A unified Nigerian Pidgin English speech-to-text dataset that combines
publicly available Pidgin ASR sources into a single train / validation /
test setup with a consistent schema. Built for fine-tuning Whisper-family
models on Nigerian Pidgin (Naija, `pcm`).

**~8.6 hours, 4,278 clips, 10 source speakers, 16 kHz mono WAV.**

Used to train [`michaelodafe/whisper-pidgin-v1`](https://huggingface.co/michaelodafe/whisper-pidgin-v1)
(21.37% WER on the test split, beating the published Wav2Vec2-XLSR-53
baseline by 8.2 pp).

## Sources

| Source | Clips | Hours | License | Notes |
|---|---|---|---|---|
| [`asr-nigerian-pidgin/nigerian-pidgin-1.0`](https://huggingface.co/datasets/asr-nigerian-pidgin/nigerian-pidgin-1.0) | 4,277 | ~8.6 h | CC-BY-4.0 | 10 native speakers (5 M / 5 F, ages 20–28), studio-quality |
| [`Rexe/nigerian-pidgin-speech`](https://huggingface.co/datasets/Rexe/nigerian-pidgin-speech) | 73 | ~0.05 h | unspecified | Single YouTube song; routed to **test only** |

The Rexe set is too small to add training signal, so it's routed
entirely to the test pool to slightly broaden the eval distribution
beyond the studio recordings.

## Splits

| Split | Clips | Hours | Mean dur | Min/Max dur |
|---|---|---|---|---|
| train | 2,708 | 5.41 | 7.2 s | 0.5 / 40.5 s |
| validation | 677 | 1.37 | 7.3 s | 0.6 / 38.2 s |
| test | 893 | 1.78 | 7.2 s | 1.4 / 44.7 s |
| **total** | **4,278** | **~8.56** | | |

Splits preserved from the upstream `asr-nigerian-pidgin/nigerian-pidgin-1.0`
release. **Speaker IDs may be shared across splits** in the source
dataset; for stricter speaker-disjoint evaluation, consult the upstream
publication.

Note: a small number of clips exceed Whisper's 30-second input window.
Filter those out (`duration <= 30.0`) when fine-tuning Whisper.

## Schema

| Column | Type | Description |
|---|---|---|
| `audio` | `Audio(sampling_rate=16000)` | Audio array, 16 kHz mono |
| `text` | `string` | Transcription, lowercased, punctuation-light |
| `source` | `string` | Origin dataset identifier |
| `duration` | `float` | Clip duration in seconds |
| `speaker_id` | `string` | Speaker identifier (may be empty for Rexe rows) |

## How to load

```python
from datasets import load_dataset

ds = load_dataset("michaelodafe/pidgin-asr-combined")
print(ds)
# DatasetDict({
#   train: Dataset({features: ['audio','text','source','duration','speaker_id'], num_rows: 2708}),
#   validation: ...,
#   test: ...
# })

example = ds["train"][0]
audio = example["audio"]["array"]   # 16kHz float32 numpy array
text  = example["text"]              # "salt di group also tok say too much salt no good"
```

## Content notes

The data is **read-style news Pidgin** — articles from BBC News Pidgin
and similar sources, read aloud in a studio setting. Lexicon is rich in:

- Pidgin function words: `dey`, `wey`, `na`, `di`, `pikin`, `pipo`,
  `tori`, `sabi`, `becos`, `neva`, `wetin`, `oga`.
- Nigerian proper nouns: politicians, states (Lagos, Anambra, Delta,
  Kogi, etc.), political parties (APC, PDP), organizations (NEMA,
  JAMB, BRT).
- Naturally code-switched English (proper nouns, loanwords, formal
  registers).

What's **not** present:
- Casual / conversational Pidgin
- Heavy code-switching with Yoruba / Igbo / Hausa
- Older speakers (training data is ages 20–28)
- Noisy real-world acoustic conditions (street, crowd, vehicle, etc.)

## Build pipeline / reproducibility

The combination, normalization, and dedupe pipeline is open source:

- Code: [michaelodafe/Naija-Pidgin-Whisper · `scripts/01_fetch_data.py`](https://github.com/michaelodafe/Naija-Pidgin-Whisper/blob/main/scripts/01_fetch_data.py)

To rebuild from sources:

```bash
git clone https://github.com/michaelodafe/Naija-Pidgin-Whisper.git
cd Naija-Pidgin-Whisper
pip install -r requirements.txt
HF_HUB_DISABLE_XET=1 python scripts/01_fetch_data.py
```

## License and attribution

This combined dataset is released under **CC-BY-4.0**, inheriting the
primary source license.

If you use this dataset, **attribution to the upstream sources is
required**:

- The bulk of the data (4,277 clips) comes from
  [`asr-nigerian-pidgin/nigerian-pidgin-1.0`](https://huggingface.co/datasets/asr-nigerian-pidgin/nigerian-pidgin-1.0)
  by the Nigerian Pidgin ASR project team. Please cite their work in
  any publication.
- A small subset (73 clips, test-only) comes from
  [`Rexe/nigerian-pidgin-speech`](https://huggingface.co/datasets/Rexe/nigerian-pidgin-speech).

## Citation

```bibtex
@misc{odafe2026pidginasrcombined,
  title  = {Pidgin ASR Combined: a unified Nigerian Pidgin speech corpus},
  author = {Odafe, Michael},
  year   = {2026},
  url    = {https://huggingface.co/datasets/michaelodafe/pidgin-asr-combined},
  note   = {Combines asr-nigerian-pidgin/nigerian-pidgin-1.0 (CC-BY-4.0) and Rexe/nigerian-pidgin-speech}
}
```

And please also cite the primary upstream source:

```bibtex
@misc{nigerianpidginasr2025,
  title  = {Nigerian Pidgin ASR Dataset v1.0},
  author = {asr-nigerian-pidgin project team},
  year   = {2025},
  url    = {https://huggingface.co/datasets/asr-nigerian-pidgin/nigerian-pidgin-1.0}
}
```

## Related

- 🤖 **Model trained on this dataset:** [michaelodafe/whisper-pidgin-v1](https://huggingface.co/michaelodafe/whisper-pidgin-v1)
- 💻 **Source code and full design notes:** https://github.com/michaelodafe/Naija-Pidgin-Whisper
- 🎤 **Live demo:** [HF Space](https://huggingface.co/spaces/michaelodafe/pidgin-whisper)