Short Description

The Graz Read and Spontaneous Speech (GRASS) Corpus, recorded at TU Graz, contains around 38 hours of spontaneous two-person conversations. Since the speakers knew each other well, the recordings are highly natural. However, such natural conversations also come with challenges: there is a large amount of overlapping speech, and in addition, audible crosstalk between the two microphones used for recording each speaker.

Each recording consists of two channels (one per speaker), but to make the corpus more usable, the goal of this project is to create cleanly separated channels where both crosstalk and overlapping speech are removed. This is particularly challenging because the crosstalk path changes dynamically as speakers move during conversation, making the task an adaptive filtering problem.

Teams of 1–2 students are welcome!

Your Tasks (depending on specific project):

• Literature research into adaptive algorithms or deep learning approaches for crosstalk and overlap removal
• Implementation of promising algorithms
• Validation and evaluation of implemented methods
• Analysis and reporting of results (thesis writing)

Your Profile

• Good knowledge of signal processing and adaptive filtering/systems
• Good programming skills in Python, MATLAB, or Julia

Contact:

Michael Paierl (paierl@tugraz.at)
Martin Hagmüller (hagmueller@tugraz.at)