Sagalee — Oromo ASR Dataset

This is a verbatim upload of the Sagalee dataset, an open-source automatic speech recognition dataset for the Oromo language, originally released by Turi Abu et al. and accepted at ICASSP 2025.

License notice: This dataset is released under CC BY-NC 4.0 — use for commercial purposes is not permitted.

Subset

Single subset orm (Oromo, ISO 639-2). No subset needed for single-language datasets.

Splits

Split	Speakers	Description
train	259	Training split
dev	12	Development split
test	8	Test split

Usage

from datasets import load_dataset

ds = load_dataset("hadamard-2/sagalee", "orm")
train = load_dataset("hadamard-2/sagalee", "orm", split="train")

Schema

Column	Type	Description
id	string	Unique utterance identifier (e.g. sagalee_SPKR014_007)
speaker_id	string	Speaker identifier (e.g. SPKR014)
transcription	string	Text transcript of the audio
language	string	ISO 639-2 language code (orm)
gender	string	Speaker gender (Male or Female)
age_group	string	Self-declared age group of the speaker (e.g. 17-24, 25-32)
dialect	string	Self-declared Oromo dialect of the speaker (e.g. Wallagga-Maccaa, Arsii-Baale)
audio	Audio	Original audio at 24kHz

Note: Column descriptions reflect our best understanding from the official dataset documentation and direct inspection. We recommend cross-referencing with the official repository for authoritative definitions.

Audio

Audio is stored as embedded binary at the original 24kHz sampling rate. No resampling has been applied. No sampling rate metadata has been declared in the parquet files, as this is a verbatim archival upload and we did not want to risk incorrect metadata.

Data Quality Notes

Four duplicate entries were found in speaker_metadata.csv:

Speaker	Issue
SPKR174	Exact duplicate row
SPKR036	Exact duplicate row
SPKR045	Exact duplicate row
SPKR032	Conflicting metadata — same speaker ID with different age_group and dialect values

For all cases the first occurrence was kept. SPKR032's conflicting metadata is a known data quality issue in the original release.

Dataset Statistics

• Total: ~100 hours of speech
• Speakers: 283 (150 Male, 133 Female)
• Collection: Crowdsourced via mobile app
• Domain: Read speech, clean and noisy environments
• Language: Afaan Oromoo (Oromo), widely spoken in Ethiopia and neighboring regions

License

CC BY-NC 4.0 International. Use for non-commercial purposes only. Full license here.

Attribution

Original dataset collected and released by Turi Abu, Ying Shi, Thomas Fang Zheng, and Dong Wang.

• Paper: Sagalee: an Open Source Automatic Speech Recognition Dataset for Oromo Language
• Source: OpenSLR 157
• GitHub: turinaf/sagalee

Citation

@INPROCEEDINGS{10890761,
  author={Abu, Turi and Shi, Ying and Zheng, Thomas Fang and Wang, Dong},
  booktitle={ICASSP 2025 - 2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
  title={Sagalee: an Open Source Automatic Speech Recognition Dataset for Oromo Language},
  year={2025},
  pages={1-5},
  doi={10.1109/ICASSP49660.2025.10890761}
}