Source code for the paper "Adversarial Data Augmentation Network for Speech Emotion Recognition" and "Improving Speech Emotion Recognition with Adversarial Data Augmentation Network".

Python 3.x
Tensorflow 1.10.0
Keras 2.1.6
numpy
tqdm
h5py
scipy
seaborn
matplotlib
scikit_learn

The emotion feature vectors were extracted using the openSMILE toolkit. Codes in folders matlab and scripts can help extract emotion features from waveforms and store the feature vectors in .mat files.

• train_adan.py trains the augmentation network based on adversarial learning (ADAN) or based on Wasserstein divergence (WADAN), and creates augmented sets;
• adan.py defines the network structure and training process of ADAN and WADAN;
• utils.py defines the additional functions used in adan.py;
• tsne_plot.py defines the functions for plotting the t-SNE plots reflecting the distributions of the real samples and synthetic samples.
• combine_h5.py defines the function to merge all augmented sets and save them into a single .h5 file using a virtual dataset.

• svm_cls.py trains an SVM classifier and outputs the UAR and WA;
• dnn_cls.py trains a DNN classifier and outputs the UAR and WA.

To train an ADAN or WADAN with default settings:

python train_adan.py 

Apart from changing the python code file to redefine the parameters, you can also define the parameters directly through the command line. For example, when you want to change the dataset used to train the model:

python train_adan.py --dataset=iemocap  

and changing the number of training epochs:

python train_adan.py --dataset=iemocap --epoch=400