KuralHub lists 4 speech emotion recognition datasets for Bengali, 4 of them openly available. Each entry below gives the emotion labels, size, recording setup, licence, where to get it and how to cite it. Bengali is also part of the benchmark. The best of 11 models reached 89.0% test accuracy (whisper-large).
Bengali 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 22 runs the mean test accuracy was 46.0% (median 39.0%). The best run reached 89.0% with whisper-large. Compare with other languages.
BANSpEmo (run ID BANSpEmo). Scores on the test split, best run highlighted.
BANSpEmo is the third Bangla emotional speech dataset for SER, containing 792 recordings of six basic emotions—Disgust, Happy, Sad, Surprised, Anger, and Fear. The dataset includes recordings from 22 unprofessional speakers (11 male, 11 female) using two sets of six sentences each, recorded in a realistic manner.
Used in the KuralHub benchmark (run ID BANSpEmo). See the results above.
Citation
@article{hussain2023banspemo,
title={Banspemo: A bangla emotional speech recognition dataset},
author={Hussain, Md Gulzar and Rahman, Mahmuda and Sultana, Babe and Shiren, Ye},
journal={arXiv preprint arXiv:2312.14020},
year={2023}
}
BanglaSER is a Bangla speech emotion recognition dataset featuring 1,467 recordings from 34 speakers (17 male, 17 female) across five emotions: angry, happy, neutral, sad, and surprise. The dataset is recorded using smartphones and laptops in real-life conditions, ensuring balanced gender distribution and compatibility with deep learning models like CNN, LSTM, and BiLSTM.
Citation
@article{das2022banglaser,
title={BanglaSER: A speech emotion recognition dataset for the Bangla language},
author={Das, Rakesh Kumar and Islam, Nahidul and Ahmed, Md Rayhan and Islam, Salekul and Shatabda, Swakkhar and Islam, AKM Muzahidul},
journal={Data in Brief},
volume={42},
pages={108091},
year={2022},
publisher={Elsevier}
}
The KUET Bangla Emotional Speech (KBES) dataset is a Bangla speech emotion recognition (SER) dataset containing 900 speech samples from 35 actors (20 females and 15 males) across diverse age ranges. The dataset includes five emotional categories: Neutral, Happy, Sad, Angry, and Disgust. Except for Neutral, each emotion is further divided into two intensity levels: Low and High, resulting in a total of nine classes. The speech samples are sourced from Bangla Telefilms, Dramas, TV Series, and Web Series, making the dataset more natural and expressive. Each class contains 100 gender-balanced samples, ensuring data symmetry for effective emotion classification.
Citation
@article{billah2023kbes,
title={KBES: A dataset for realistic Bangla speech emotion recognition with intensity level},
author={Billah, Md Masum and Sarker, Md Likhon and Akhand, MAH},
journal={Data in Brief},
volume={51},
pages={109741},
year={2023},
publisher={Elsevier}
}
SUBESCO is a Bangla emotional speech corpus containing 7,000 sentence-level utterances recorded by 20 professional actors (10 male, 10 female) across seven emotions: Anger, Disgust, Fear, Happiness, Neutral, Sadness, and Surprise. The dataset spans 7 hours and 40 minutes, totaling 2.03 GB. Evaluated by 50 raters, it achieved a human perception accuracy of 71%. Details on its creation and analysis are available in its PLOS ONE journal publication.
Used in the KuralHub benchmark (run ID SUBESCO). See the results above.
Citation
@article{sultana2021sust,
title={SUST Bangla Emotional Speech Corpus (SUBESCO): An audio-only emotional speech corpus for Bangla},
author={Sultana, Sadia and Rahman, M Shahidur and Selim, M Reza and Iqbal, M Zafar},
journal={Plos one},
volume={16},
number={4},
pages={e0250173},
year={2021},
publisher={Public Library of Science San Francisco, CA USA}
}
Know another Bengali speech emotion dataset, or a newer version of one listed here? Suggest it or open a pull request.
This page is generated from datasets/Bengali/ in the KuralHub repository.
If you use a dataset, cite its authors. If KuralHub helped you find it, please also cite the KuralHub paper.