This model is a Perceiver-based (https://huggingface.co/docs/transformers/model_doc/perceiver) emotion recognition model trained on RAVDESS dataset (https://zenodo.org/record/1188976#.Y5iqPy2B1QI). The model is trained using 3 modalities: video, audio, and text.

For details on the data collection, check here: https://zenodo.org/record/1188976

The feature extraction for each modality and training procedure follows the steps mentioned here: https://dl.acm.org/doi/10.1145/3551876.3554806

You can use the raw model for directly recognize emotion (classes: 01 = neutral, 02 = calm, 03 = happy, 04 = sad, 05 = angry, 06 = fearful, 07 = disgust, 08 = surprised) or fine-tune on a downstream task.

The model is trained on only one dataset and uses 8 specific classes of emotions. The limitation lies in the lack of diversity in the demographics and emotions.