Harcuracy/google_waxal_asr_challenge

WaxalNLP ASR — Cleaned Subset (Lingala, Shona, Luganda)

This dataset is a cleaned, corrected subset of google/WaxalNLP, covering the train and validation splits for three languages:

• —lin_asr — Lingala
• —sna_asr — Shona
• —lug_asr — Luganda

The test split from the original dataset is intentionally excluded.

What was changed

The original transcriptions for these three languages contained a number of errors. A corrected transcription file was applied on top of the original audio, matched by example id. Corrections were merged in as follows:

Coverage was effectively complete: every row in every split received a corrected transcription, with the exception of a single lin train example whose id had no corresponding entry in the correction file — that row was dropped rather than kept with a potentially unverified transcription.

Audio is unchanged from the original WaxalNLP release and is embedded directly in this dataset (not referenced by external path), so no separate audio download is required.

Fields

Each example contains:

• —id — unique example identifier
• —speaker_id — anonymized speaker identifier
• —transcription — corrected transcription text
• —language — language code (lin, sna, or lug)
• —gender — speaker gender
• —audio — embedded audio (bytes + original path)

Usage

from datasets import load_dataset

# Load a specific language
lin = load_dataset("Harcuracy/google_waxal_asr_challenge", "lin_asr")
sna = load_dataset("Harcuracy/google_waxal_asr_challenge", "sna_asr")
lug = load_dataset("Harcuracy/google_waxal_asr_challenge", "lug_asr")

train = lin["train"]
val = lin["validation"]

example = train[0]
print(example["transcription"])
print(example["audio"])  # {'bytes': ..., 'path': ...}

Source and license

This dataset is derived from google/WaxalNLP, released by Google under the CC-BY-4.0 license. The original Waxal project collected ASR and TTS data for African languages in partnership with Makerere University, the University of Ghana, Digital Umuganda, Media Trust, Loud and Clear, and AIMS Senegal, with funding from Google and the Gates Foundation.

This derivative work is released under the same CC-BY-4.0 license. Please cite the original WaxalNLP paper and dataset when using this data.

Limitations

• —Only train and validation splits are included; no test split is provided here.
• —Transcription corrections were applied based on a single external correction pass and have not been independently re-verified beyond automated id/split matching.
• —Only three of WaxalNLP's languages are covered (Lingala, Shona, Luganda).