Prosody-Aware Graph Neural Networks for Speech Emotion Recognition

Create subgraphs based on the prosodic content of the utterances

Execute the following commands to generate the prosodic patterns for train and test sets:

python preprocess.py -a '~/IEMOCAP/train/'

python preprocess.py -a '~/IEMOCAP/test/'

Generate graph with the following command:

python preprocess.py -g '~/IEMOCAP/train/'
python preprocess.py -g '~/IEMOCAP/test/'

The IEMOCAP /train directory should follow this structure:

• ang
• sad
• hap
• neu

Where each emotion label is a subdirectory of the train/ directory.

Pretrain the speech model

Use the following command where -b and -e parameters are batch size and epochs respectively.

python main.py -ptrain '~/IEMOCAP/train/' -ptest '~/IEMOCAP/test/' -b 128 -e 100

Train and test the graph model

python main.py -d 'patterns/' -b 64 -e 200

where patterns/ is the directory where the prosodic patterns are located.