Arabic speech emotion recognition datasets

KuralHub lists 2 speech emotion recognition datasets for Arabic, 2 of them openly available. Each entry below gives the emotion labels, size, recording setup, licence, where to get it and how to cite it. Arabic is also part of the benchmark. The best of 11 models reached 82.0% test accuracy (whisper-large).

Language family: Afro-Asiatic (Semitic)
ISO code: ar
Spoken mainly in: Arab World, Middle East and North Africa
Datasets listed: 2
In benchmark: Yes

Datasets at a glance

Dataset	Type	Samples	Speakers	Recording	Access	Licence
ANAD	Audio	1384		Spontaneous	Open	CC BY-NC-SA 4.0
BAVED	Audio	1935	61	Acted	Open

Benchmark results

Arabic 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 66.1% (median 64.0%). The best run reached 82.0% with whisper-large. Compare with other languages.

**ANAD** (run ID `ANAD`). Scores on the test split, best run highlighted.
Model	Family	Val acc.	Test acc.	Macro-F1	Weighted-F1	Logs
hubert-base-ls960	HuBERT	0.63	0.58	0.58	0.58	logs
hubert-large-ls960-ft	HuBERT	0.64	0.66	0.51	0.59	logs
wav2vec2-base	wav2vec 2.0	0.62	0.57	0.45	0.48	logs
wav2vec2-large-960h	wav2vec 2.0	0.62	0.65	0.54	0.61	logs
wav2vec2-large-lv60	wav2vec 2.0	0.64	0.64	0.39	0.51	logs
wav2vec2-xls-r-300m	wav2vec 2.0	0.61	0.61	0.46	0.54	logs
wav2vec2-xls-r-1b	wav2vec 2.0	0.61	0.63	0.47	0.56	logs
wavlm-base-plus	WavLM	0.66	0.61	0.54	0.57	logs
wavlm-large	WavLM	0.66	0.71	0.57	0.64	logs
whisper-small	Whisper	0.72	0.79	0.78	0.79	logs
whisper-largebest	Whisper	0.78	0.82	0.81	0.81	logs

ANAD

Arabic Natural Audio Dataset

Type: Audio
Emotion labels: Happy, Angry, Surprised
Size: Number of samples: 1384
Recording: Spontaneous
Access: Open
Licence: CC BY-NC-SA 4.0
File format: .wav
Website: Kaggle
Download: Kaggle

The Arabic Natural Audio Dataset (ANAD) was created to recognize emotions in speech, specifically targeting the detection of happiness, anger, and surprise in Arabic conversations. It aims to enhance communication for the deaf community by embedding emotion detection into speech recognition systems, improving their ability to interpret both speech and emotional context. The dataset was collected from eight videos of live calls from Arabic talk shows, with speech segments manually labeled by 18 listeners for emotional content. Each video was divided into 1-second speech units, resulting in a corpus of 1,384 records. Acoustic features, such as MFCC, intensity, and fundamental frequency, were extracted and analyzed using 19 statistical functions, totaling 950 features. This dataset is a step toward improving speech-to-text and text-to-speech systems for the hearing-impaired.

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

Citation

@misc{arabic_natural_audio_dataset,
  title = {Arabic Natural Audio Dataset},
  author = {Samira Klaylat},
  year = {2022},
  publisher = {Kaggle},
  howpublished = {\url{https://www.kaggle.com/datasets/suso172/arabic-natural-audio-dataset}},
  note = {Accessed: 2025-05-10}
}

TBD-Add kaggle link as reference

Source: datasets/Arabic/ANAD.md · Suggest an edit

BAVED

Basic Arabic Vocal Emotions Dataset

Type: Audio
Emotion labels: Low, Neutral, High
Size: Number of samples 1935, speakers: 61
Speakers: 61
Recording: Acted
Access: Open
Licence: Not stated
File format: .wav
Download: GitHub

Basic Arabic Vocal Emotions Dataset (BAVED) is a dataset that contains an arabic words spelled in diffrent levels of expressed emotions recorded in an audio/wav format.

Citation

@misc{aouf2019baved,
  title        = {Basic Arabic Vocal Emotions Dataset (BAVED)},
  author       = {Ali Aouf},
  year         = {2019},
  publisher    = {GitHub},
  howpublished = {\url{https://github.com/40uf411/Basic-Arabic-Vocal-Emotions-Dataset}},
  note         = {Accessed: 2025-05-10}
}

TBD - Need to add github as reference

Source: datasets/Arabic/BAVED.md · Suggest an edit

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

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