Cross-Lingual Cross-Age Group Adaptation for Low-Resource Elderly Speech Emotion Recognition

In this work, we study the prospect of transferring emotion recognition ability over various age groups and languages through the utilization of multilingual pre-trained speech models. In this work, we develop two speech emotion recognition resources, i.e., BiMotion and YueMotion.

• BiMotion is a bi-lingual bi-age-group speech emotion recognition benchmark that covers 6 adult and elderly speech emotion recognition datasets from English and Mandarin Chinese.
• YueMotion is a newly-constructed publicly-available Cantonese speech emotion recognition dataset.

Dataset

YueMotion, which is the Cantonese speech emotion recognition dataset we collect, is available on HuggingFace.

Fine-tuned Model Checkpoints

• English-Chinese all age: HuggingFace
• English-Chinese elderly: HuggingFace
• English-Chinese adults: HuggingFace

Publication

Our paper will be published at INTERSPEECH 2023. In the meantime, you can find our paper on arXiv. If you find our work useful, please consider citing our paper as follows:

@misc{cahyawijaya2023crosslingual,
      title={Cross-Lingual Cross-Age Group Adaptation for Low-Resource Elderly Speech Emotion Recognition}, 
      author={Samuel Cahyawijaya and Holy Lovenia and Willy Chung and Rita Frieske and Zihan Liu and Pascale Fung},
      year={2023},
      eprint={2306.14517},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}