KuralHub lists 2 speech emotion recognition datasets for Portuguese, 1 of them openly available. Each entry below gives the emotion labels, size, recording setup, licence, where to get it and how to cite it. Portuguese is also part of the benchmark. The best of 11 models reached 91.0% test accuracy (whisper-small).
Portuguese 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 60.0% (median 63.0%). The best run reached 91.0% with whisper-small. Compare with other languages.
emoUERJ (run ID emoUERJ). Scores on the test split, best run highlighted.
The VERBO database is a speech emotion dataset designed to evaluate the effectiveness of Speech Emotion Recognition (SER) systems. It comprises recordings of 14 sentences spoken by 12 professional speakers, including six actors and six actresses. Each speaker expresses six basic emotions—happiness, sadness, anger, fear, disgust, and surprise—along with a neutral state. In total, the dataset contains 1,176 audio recordings. A statistical analysis was conducted to assess how well these recordings represent emotional speech, making the VERBO database a valuable resource for training and evaluating SER models.
Citation
@article{torres2018verbo,
title={Verbo: voice emotion recognition database in Portuguese language},
author={Torres Neto, Jos{\'e} R and Mano, Leandro Y and Ueyama, J{\'o} and others},
journal={Journal of Computer Science},
volume={14},
number={11},
pages={1420--1430},
year={2018}
}
The dataset was developed at the State University of Rio de Janeiro to address the challenge of language specificity in Speech Emotion Recognition (SER), particularly for Portuguese, where existing databases are limited. It consists of 377 audio recordings, each featuring an actor expressing one of four emotions: happiness (91 samples), anger (94 samples), sadness (100 samples), and neutrality (92 samples). Each file is named using a structured format that encodes the actor's gender, ID, expressed emotion, and recording number. For example, "w04a11" represents the eleventh recording of actress 04 portraying anger. This dataset aims to facilitate the development of SER models tailored to the Portuguese language.
Used in the KuralHub benchmark (run ID emoUERJ). See the results above.
Citation
@article{bastos2021emouerj,
title={emoUERJ: an emotional speech database in Portuguese},
author={Bastos Germano, RG and Pompeu Tcheou, M and da Rocha Henriques, F and Pinto Gomes Jr, S},
journal={Zenodo. Sep},
year={2021}
}
Know another Portuguese speech emotion dataset, or a newer version of one listed here? Suggest it or open a pull request.
This page is generated from datasets/Portuguese/ in the KuralHub repository.
If you use a dataset, cite its authors. If KuralHub helped you find it, please also cite the KuralHub paper.