Persian speech emotion recognition datasets

KuralHub lists 1 speech emotion recognition dataset for Persian. Each entry below gives the emotion labels, size, recording setup, licence, where to get it and how to cite it. Persian is also part of the benchmark. The best of 11 models reached 83.0% test accuracy (whisper-small).

Language family: Indo-European (Iranian)
ISO code: fa
Spoken mainly in: Iran, Afghanistan, and Tajikistan
Datasets listed: 1
In benchmark: Yes

Datasets at a glance

Dataset	Type	Samples	Speakers	Recording	Access	Licence
ShEMO	Audio	3000	87	Acted	Open

Benchmark results

Persian 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 67.1% (median 76.0%). The best run reached 83.0% with whisper-small. Compare with other languages.

**ShEMO** (run ID `SHEMO`). Scores on the test split, best run highlighted.
Model	Family	Val acc.	Test acc.	Macro-F1	Weighted-F1	Logs
hubert-base-ls960	HuBERT	0.77	0.80	0.54	0.78	logs
hubert-large-ls960-ft	HuBERT	0.57	0.58	0.31	0.53	logs
wav2vec2-base	wav2vec 2.0	0.71	0.76	0.46	0.73	logs
wav2vec2-large-960h	wav2vec 2.0	0.44	0.44	0.18	0.33	logs
wav2vec2-large-lv60	wav2vec 2.0	0.43	0.43	0.16	0.30	logs
wav2vec2-xls-r-300m	wav2vec 2.0	0.44	0.46	0.19	0.35	logs
wav2vec2-xls-r-1b	wav2vec 2.0	0.60	0.67	0.36	0.63	logs
wavlm-base-plus	WavLM	0.77	0.79	0.52	0.77	logs
wavlm-large	WavLM	0.72	0.80	0.59	0.79	logs
whisper-smallbest	Whisper	0.79	0.83	0.62	0.82	logs
whisper-large	Whisper	0.80	0.82	0.61	0.81	logs

ShEMO

Sharif Emotional Speech Database

Type: Audio
Emotion labels: Anger, Fear, Happiness, Sadness, Surprise, Neutral
Size: samples: 3000, speakers: 87, duration: 3 h and 25 min
Speakers: 87
Recording: Acted
Access: Open
Licence: Not stated
File format: .wav
Paper: Paper Link
Download: Download

ShEMO is a large-scale Persian emotional speech database containing 3,000 semi-natural utterances (3 hours 25 minutes) from 87 native speakers. It includes six emotion categories: anger, fear, happiness, sadness, surprise, and neutral. The dataset was annotated by 12 raters with a 64% inter-annotator agreement and is available for academic research.

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

Citation

@article{mohamad2019shemo,
  title={ShEMO: a large-scale validated database for Persian speech emotion detection},
  author={Mohamad Nezami, Omid and Jamshid Lou, Paria and Karami, Mansoureh},
  journal={Language Resources and Evaluation},
  volume={53},
  pages={1--16},
  year={2019},
  publisher={Springer}
}

Source: datasets/Persian/ShEMO.md · Suggest an edit

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

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