Link to article: here
To run our code, choose between obtaining the dataset from this link or accessing our preprocessed data here. For the processed data, unzip and save it in the data folder after download. If opting for the complete IEMOCAP data, unzip it and process using the file here(scripts/data collected.ipynb) or use !python mocap_data_collect.py.
The dataset includes around ten emotions, but our study focuses on four: anger, neutral, excitement, and sadness. The processed data file exclusively contains these emotions. If you want more detected emotions, process the data independently.
After data processing:
-
Run any of these files (Speech, TextBERT, MotionCap)
- Speech_Wav2vec: for training on speech only. Find fine-tuning details here.
Speech data was processed separately.
- TextBERT: for training on text.
- MotionCap: for training on motion capture (Mocap_Rotate_Adam or SGD).
- Concatenation: used to train all three modalities (speech + text + motion capture).
-
Additional files can be run for comparison purposes (Pairwise_Concat and Ablation):
- Pairwise_Concatenation: combines two modalities instead of three.
- Ablation: includes concatenation files trained using different optimizers.
The code can be executed on Google Colab, Kaggle, or create your own environment on your machine.
The files most pertinent...
.
├── ...
├── codes
│ ├── Ablation
│ ├── Speech_Wav2vec
| ├── TextBERT
| ├── Mocap_Rotate_...
│ └── Concatenation_...
├── data
└── Readme
Please refer to our article for more details.
@misc{patamia2023multimodal, title={Multimodal Speech Emotion Recognition Using Modality-specific Self-Supervised Frameworks}, author={Rutherford Agbeshi Patamia and Paulo E. Santos and Kingsley Nketia Acheampong and Favour Ekong and Kwabena Sarpong and She Kun}, year={2023} }