On Enhancing Speech Emotion Recognition using Generative Adversarial Networks
This work addresses the challenge of limited labeled data for speech emotion recognition, but it is incremental as it applies existing GAN methods to a specific domain.
The paper tackles the problem of improving speech emotion recognition by using Generative Adversarial Networks (GANs) to generate synthetic feature vectors, and finds that this approach enhances classifier performance in cross-validation and cross-corpus studies.
Generative Adversarial Networks (GANs) have gained a lot of attention from machine learning community due to their ability to learn and mimic an input data distribution. GANs consist of a discriminator and a generator working in tandem playing a min-max game to learn a target underlying data distribution; when fed with data-points sampled from a simpler distribution (like uniform or Gaussian distribution). Once trained, they allow synthetic generation of examples sampled from the target distribution. We investigate the application of GANs to generate synthetic feature vectors used for speech emotion recognition. Specifically, we investigate two set ups: (i) a vanilla GAN that learns the distribution of a lower dimensional representation of the actual higher dimensional feature vector and, (ii) a conditional GAN that learns the distribution of the higher dimensional feature vectors conditioned on the labels or the emotional class to which it belongs. As a potential practical application of these synthetically generated samples, we measure any improvement in a classifier's performance when the synthetic data is used along with real data for training. We perform cross-validation analyses followed by a cross-corpus study.