KuralHub lists 2 speech emotion recognition datasets for Japanese, 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. Japanese is also part of the benchmark. The best of 11 models reached 87.0% test accuracy (whisper-large).
Japanese 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 61.9% (median 67.0%). The best run reached 87.0% with whisper-large. Compare with other languages.
JVNV (run ID JVNV). Scores on the test split, best run highlighted.
The JVNV dataset is a Japanese emotional speech corpus that uniquely integrates both verbal expressions and nonverbal vocalizations, such as laughter and sobbing, to capture a richer emotional spectrum. It contains 3.94 hours of high-quality speech recorded at 48 kHz in WAV format from four speakers (two male, two female), covering six emotions: anger, disgust, fear, happiness, sadness, and surprise. The dataset comprises 1,615 utterances, each expressing a specific emotion, with annotated time intervals for nonverbal sounds. JVNV also features two recording conditions: a regular session, where nonverbal phrases are predefined, and a phrase-free session, where speakers decide their own expressions. Notably, the dataset’s phoneme-balanced scripts were generated using a large language model, making JVNV the first emotional speech corpus to utilize AI-generated content. With a 94% subjective emotion recognition accuracy—surpassing previous Japanese emotional speech datasets—JVNV stands as a valuable resource for speech emotion recognition research.
Used in the KuralHub benchmark (run ID JVNV). See the results above.
Citation
@article{xin2024jvnv,
title={JVNV: A corpus of japanese emotional speech with verbal content and nonverbal expressions},
author={Xin, Detai and Jiang, Junfeng and Takamichi, Shinnosuke and Saito, Yuki and Aizawa, Akiko and Saruwatari, Hiroshi},
journal={IEEE Access},
volume={12},
pages={19752--19764},
year={2024},
publisher={IEEE}
}
This dataset contains human and artificial speech samples with emotional expressions. The human speech was recorded from a single Japanese male speaker in a soundproof room, while the artificial speech was synthesized using a subset of the dataset for training. It includes words with different syllables and accent types, covering 47 distinct emotions such as anger, joy, worry, and relief. The audio files are in WAV format (16 kHz, 16-bit, mono).
Citation
@article{moriyama2011keio,
title={Keio University Japanese Emotional Speech Database (Keio-ESD)},
author={Moriyama, T},
journal={Speech Resources Consortium, National Institute of Informatics,(dataset)},
year={2011}
}
Know another Japanese speech emotion dataset, or a newer version of one listed here? Suggest it or open a pull request.
This page is generated from datasets/Japanese/ in the KuralHub repository.
If you use a dataset, cite its authors. If KuralHub helped you find it, please also cite the KuralHub paper.