In recent years Generative Adversarial Neural Networks (GANs) have caused a lot of excitement in the computer vision community. They have proven to be an outstanding tool for producing artificial images, which are indistinguishable from real images. Additionally, the introduction of the Wasserstein loss has made the training of GANs easier, while achieving improved results. In this project we would like to explore the capabilities of Wasserstein GANs in the realm of time-series analysis. On the example of an audio source-separation task, we would like to investigate the use of GANs to augment the training data, while also using the generative part of the model to create realistic artificial audio samples.

Objectives:

• Implementation of a Wasserstein GAN framework, applicable to audio data
• Evaluation of a GAN for data augmentation in a standard source-separation task
• Generation of artificial audio samples

Your Profile:

• Master student in telematics/electrical engineering/similar
• Programming skills in Python or R
• Familiarity with one of the large Machine Learning frameworks (Tensorflow/Pytorch)
• Affinity for audio processing advantageous