IEEE SPCup'22 - 5th Place

Oct 8, 2023   | #Deep Learning  #Signal Processing  #Adversarial Learning  

Competition Title	Signal Processing Cup 2022
Organizer	IEEE International Organization
Year	Aug 2022
Place	5th place
Team	Nuwan, Oshan, Sachini, Sandani, Chalani, Dilmi, Praneeth, Isuru
Code	Paper

Description

In the challenging domain of SPCup'22, our mission was to distinguish synthetically generated human voices from authentic human voice clips, a critical task with far-reaching implications. Notably, we dealt with voice signals heavily processed by effects such as reverb, dry-wet reflection, and echoing, making the distinction even more intricate. Our journey began with the implementation of a transformer model for an initial classification of these complex signals. As we delved deeper into the project, we harnessed the power of Fast Fourier Transformation-based CNN modeling to further enhance our model's performance, ensuring that the line between authentic and synthetic human voices was crystal clear.

Technologies

Methodologies

Transformer Model Fast Fourier Transformation (FFT) Contrastive Learning

Key Achievements

• Effective Signal Classification: We successfully employed a transformer model for the initial classification of voice signals, providing a solid foundation for our project.
• Robust Signal Processing: Dealing with heavily processed audio signals underscored the importance of robust signal processing techniques.
• Model Versatility: Combining transformer models with FFT and CNNs illustrated the power of diverse methodologies in tackling complex audio classification tasks.