Kolokwa: Liberian English Speech Dataset

Version: v2026-Q2

The first open voice dataset for Liberian English (Kolokwa), a creole spoken by approximately 5 million people in Liberia and the global diaspora. Built through crowdsourced contributions on the Kolokwa platform by the Palava Hut Foundation.

Dataset Summary

• Total recordings: 87
• Total duration: 6.1 minutes
• Language: Liberian English (Kolokwa)
• License: CC BY-NC-SA 4.0
• Layout: Three Parquet tables (recordings.parquet, sentences.parquet, consent_records.parquet) at the snapshot root, plus a flat audio/ directory and the literal CONSENT_TEXT file.

Splits

The split assignment is recorded as a column inside recordings.parquet (no per-split file tree). Splits are deterministic per category at an 80 / 10 / 10 ratio.

Split	Recordings
train	70
test	9
validation	8

Categories

Category	Recordings
exclamations	33
requests	22
greetings	21
numbers_and_time	5
directions	5
daily_life	1

Audio format

Native browser-captured format (typically WebM/Opus at 48 kHz). Consumers should resample as needed; Whisper and most modern ASR toolchains do this internally.

The format column in recordings.parquet carries the bare container codec (webm, wav, ogg, ...). The sample_rate column is intentionally null for every row because the database does not store the true capture rate — consumers must decode the audio header for the true rate (e.g. librosa.load(path, sr=None) or ffprobe). Passing sr=16000 to librosa.load is the recommended path for ASR.

Data fields

recordings.parquet

Column	Type	Description
recording_id	string	Unique recording identifier
sentence_id	string	Foreign key into sentences.parquet
audio_path	string	Snapshot-relative path: audio/<recording_id>.<ext>
format	string	Bare container codec (webm, wav, ogg, ...)
sample_rate	int (nullable)	Always null (see "Audio format" above)
duration_s	float	Recording duration in seconds
content_hash	string (nullable)	SHA-256 of the audio bytes on disk
consent_id	string	Foreign key into consent_records.parquet
session_id	string	Anonymous session identifier
status	string	Always approved for published rows
split	string	One of train, test, validation
created_at	string	ISO 8601 timestamp

sentences.parquet

Column	Type	Description
sentence_id	string	Unique sentence identifier
text_kolokwa	string	Kolokwa sentence text
text_english	string	English translation
category	string	Sentence category (greetings, food, ...)
difficulty	int	Difficulty level (1-3)
source	string	Provenance label (e.g. community)

consent_records.parquet

Column	Type	Description
id	string	Consent record identifier (referenced by recordings.consent_id)
consent_text_hash	string	SHA-256 of CONSENT_TEXT at consent time
consent_version	string	Consent text version (e.g. 1.1.0)

Eval contract

This dataset is consumed by the phelinki/kolokwa-eval project, which performs WER and safety evaluation against ASR systems.

An external evaluation report (EVAL_REPORT.md) will be PR'd back to this dataset card.

The provenance audit on the eval side checks that sha256(CONSENT_TEXT) matches every row's consent_records.consent_text_hash — see the CONSENT_TEXT file at the snapshot root.

Intended Uses

This dataset is designed for:

• Training automatic speech recognition (ASR) models for Liberian English
• Fine-tuning existing models (Whisper, wav2vec2, HuBERT) on Kolokwa
• Linguistic research on Liberian English phonetics and prosody
• Building voice assistants and speech tools for Liberian speakers

Collection Process

Recordings are crowdsourced through the Kolokwa web platform. Contributors read Kolokwa sentences aloud from any device. All contributions require explicit consent with GDPR-aligned terms. Recordings are reviewed for quality before inclusion.

Licensing

This dataset is released under CC BY-NC-SA 4.0 (Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International).

Citation

If you use this dataset, please cite:

@dataset{kolokwa_v2026_Q2,
  title={Kolokwa: Liberian English Speech Dataset},
  author={Palava Hut Foundation},
  year={2026},
  version={v2026-Q2},
  url={https://huggingface.co/datasets/phelinki/kolokwa},
  license={CC BY-NC-SA 4.0}
}