Amharic speech emotion recognition datasets

KuralHub lists 1 speech emotion recognition dataset for Amharic. Each entry below gives the emotion labels, size, recording setup, licence, where to get it and how to cite it. Amharic is also part of the benchmark. The best of 11 models reached 90.0% test accuracy (whisper-small).

Language family: Afro-Asiatic (Semitic)
ISO code: am
Spoken mainly in: Ethiopia
Datasets listed: 1
In benchmark: Yes

Datasets at a glance

Dataset	Type	Samples	Speakers	Recording	Access	Licence
Amharic Speech Emotion Dataset (ASED)	Audio	2,474	65	Acted	Open

Benchmark results

Amharic is one of the 29 languages in the KuralHub benchmark. Each of the 11 frozen speech models was trained with a small classification head on the shared emotions and scored on a held-out test split. Over all 11 runs the mean test accuracy was 63.6% (median 75.0%). The best run reached 90.0% with whisper-small. Compare with other languages.

**Amharic Speech Emotion Dataset (ASED)** (run ID `ASED`). Scores on the test split, best run highlighted.
Model	Family	Val acc.	Test acc.	Macro-F1	Weighted-F1	Logs
hubert-base-ls960	HuBERT	0.88	0.89	0.89	0.89	logs
hubert-large-ls960-ft	HuBERT	0.64	0.63	0.63	0.63	logs
wav2vec2-base	wav2vec 2.0	0.75	0.75	0.75	0.75	logs
wav2vec2-large-960h	wav2vec 2.0	0.30	0.32	0.27	0.26	logs
wav2vec2-large-lv60	wav2vec 2.0	0.23	0.27	0.25	0.24	logs
wav2vec2-xls-r-300m	wav2vec 2.0	0.31	0.32	0.29	0.30	logs
wav2vec2-xls-r-1b	wav2vec 2.0	0.35	0.40	0.32	0.33	logs
wavlm-base-plus	WavLM	0.87	0.84	0.84	0.84	logs
wavlm-large	WavLM	0.80	0.82	0.82	0.82	logs
whisper-smallbest	Whisper	0.91	0.90	0.90	0.90	logs
whisper-large	Whisper	0.88	0.86	0.86	0.86	logs

Amharic Speech Emotion Dataset (ASED)

Type: Audio
Emotion labels: Neutral, Fearful, Happy, Sad, Angry
Size: 2,474 audio samples; 65 speakers (25 female, 40 male); each sample 2-4 seconds
Speakers: 65
Recording: Acted
Access: Open
Licence: Not stated
File format: .wav
Paper: A New Amharic Speech Emotion Dataset and Classification Benchmark
Download: GitHub Repository

The Amharic Speech Emotion Dataset (ASED) is a pioneering resource for Speech Emotion Recognition (SER) in the Amharic language. It encompasses recordings from 65 native speakers across four dialects: Gojjam, Wollo, Shewa, and Gonder. Participants delivered 2,474 audio samples, each lasting between 2 to 4 seconds, expressing five distinct emotions: neutral, fearful, happy, sad, and angry. The recordings underwent evaluation by eight judges, achieving a high agreement level (Fleiss kappa = 0.8). This dataset serves as a valuable benchmark for developing and testing SER models tailored to Amharic. (arxiv.org)

Note: The dataset is publicly available, so no request is needed.

Used in the KuralHub benchmark (run ID ASED). See the results above.

Citation

@article{retta2023new,
  title={A new Amharic speech emotion dataset and classification benchmark},
  author={Retta, Ephrem Afele and Almekhlafi, Eiad and Sutcliffe, Richard and Mhamed, Mustafa and Ali, Haider and Feng, Jun},
  journal={ACM Transactions on Asian and Low-Resource Language Information Processing},
  volume={22},
  number={1},
  pages={1--22},
  year={2023},
  publisher={ACM New York, NY}
}

Source: datasets/Amharic/ASED.md · Suggest an edit

Know another Amharic speech emotion dataset, or a newer version of one listed here? Suggest it or open a pull request.

This page is generated from datasets/Amharic/ in the KuralHub repository. If you use a dataset, cite its authors. If KuralHub helped you find it, please also cite the KuralHub paper.