Short Description

In principle, neural networks can learn any desired task, provided that sufficient training data is available. The currently best performance in the MNIST digit classification task was achieved by a large scale neural network, where input data is artificially altered in each training epoch.

For this project, we plan to apply large scale neural networks for the audio source separation task, where we can also generate a practically unlimited amount of training data.

Your Tasks

• Get familiar with neural networks
• Get familiar with source separation via time frequency masking
• Compare several neural network architectures for source separation
• Write a report (around 10 pages)

Your Profile/Requirements

This project is suited for Master students in Telematics, Audio Engineering, Electrical Engineering, Computer Science and Software Development.

• Interest in machine learning techniques.
• Good Matlab skills

Contact

Robert Peharz (robert.peharz@tugraz.at or 0316/873 4482)

Michael Wohlmayr (michael.wohlmayr@tugraz.at or 0316/873 4366))

References

[1] C. Bishop, “Pattern Recognition and Machine Learning,” Springer, 2006.