Sagalee – Automatic Speech Recognition Dataset for Afaan Oromoo

Dataset Description

Sagalee is Speech Recognition Dataset for Oromo language Presented in the paper:
Sagalee: an Open Source Automatic Speech Recognition Dataset for Oromo Language at ICASSP 2025

• Training Code: turinaaf/Sagalee
• Arxiv: https://arxiv.org/abs/2502.00421

The dataset is released under Attribution–NonCommercial 4.0 International (CC BY-NC 4.0) It contains read-speech recordings from native Oromo speakers (283 speakers) representing multiple dialectal regions. Each audio file is paired with its corresponding orthographic transcription. Total duration of the dataset is 100 hours.

Dataset Structure

Splits

Total duration: 100hrs

Split	Description
train	Training set 93.6hrs
dev	Development / val set 4.2hrs
test	Test set 2.4hrs

Data Fields

Field	Type	Description
audio	Audio	Audio waveform (WAV, mono) 24kHz
transcription	string	Orthographic transcription
speaker_id	string	Unique speaker identifier (e.g. SPKR001)
gender	string	Speaker gender (Male / Female)
age_group	string	Speaker age group (e.g. 17-24, 25-32)
dialect	string	Dialectal region of the speaker
split	string	Dataset split name
utterance_id	string	Unique utterance file stem

Each speaker folder (named after speaker_id) contains utterences (.wav) and transcripts (.txt) files with same name.

Dialect Regions

The following Oromo dialectal varieties are represented. Majority of utterences are from Wallagaa-Maccaa Giddugala/Shawaa-Tuulama and Arsii-Baalee

• Wallagga-Maccaa (48.8%)
• Giddugala / Shawaa-Tuulama (21.9%)
• Arsii-Baale (20.5%)
• Jimmaa-Iluu (4.2%)
• Harargee
• Boorana-Gabra
• Gujii
• Walloo-Raayya

The remaining 4 regions are 4.6% of the dataset.

Speaker Demographics

Category	Values
Gender	Male, Female (Gender is balanced)
Age Groups	17-24, 25-32, 33-40, 41-48 (Majority are 17-24 age group)

Baseline results

The dataset is used to train conformer model from scratch and finetuned the Whisper-large-v3. The following table shows baseline result evaluated on test set of Sagalee without using Language Model using Word Error Rate as metric.

Model	Training Mode	WER (%)
Conformer (AED)	From scratch	15.32
Conformer (CTC)	From scratch	18.74
Whisper Large-v3	Fine-tuning	10.82

Usage

from datasets import load_dataset

sagalee_ds = load_dataset("turiabu/Sagalee")
print(sagalee_ds)

Output:

DatasetDict({
    test: Dataset({
        features: ['audio', 'transcription', 'speaker_id', 'gender', 'age_group', 'dialect', 'split', 'utterance_id'],
        num_rows: 1178
    })
    dev: Dataset({
        features: ['audio', 'transcription', 'speaker_id', 'gender', 'age_group', 'dialect', 'split', 'utterance_id'],
        num_rows: 2387
    })
    train: Dataset({
        features: ['audio', 'transcription', 'speaker_id', 'gender', 'age_group', 'dialect', 'split', 'utterance_id'],
        num_rows: 50008
    })
})

Supported Tasks

• Automatic Speech Recognition (ASR)
• Text-to-Speech
• Dialect / Speaker Identification

Citation

If you use this dataset, please cite:

@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}
}

License

This dataset is released under the Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) license.

Contact

For questions or issues, please open a discussion on the dataset repository.