Welcome to ASPIRE
Audio & Signal Processing Research at Ghent University
Our Philosophy
We are passionate about developing robust algorithms and practical solutions for real-world applications. Our research spans speech enhancement, audio analytics, spatial audio, electrophysiological signals and biomedical signal analysis. By integrating domain knowledge with state-of-the-art machine learning, we strive to push the boundaries of what is possible in signal processing.
In addition to application-driven research, we also pursue more fundamental investigations that focus on proof-of-concept work for novel ideas. While these projects may not be immediately targeted at real-world deployment, they are inspired by genuine gaps in current knowledge and are motivated by challenges observed in practical scenarios.
Collaboration & Innovation
ASPIRE actively collaborates with both academic and industry partners. We are currently working together with industry on a one-on-one basis as well as through consortium projects and have had close collaboration with other universities in the form of exchange students and joint publications.
We welcome collaborative research with other academic partners for joint research as well as with companies seeking expertise in signal processing, machine learning, and audio technologies. If you are interested in partnering with us, please get in touch.
Students & Researchers
ASPIRE encourages individuals from all backgrounds and experience levels to collaborate with our multidisciplinary team. We invite you to explore our website, learn more about our researchers, and engage with our ongoing projects. Whether you're a bachelor's or master's student, a fellow researcher, or a potential new ASPIRE member, we welcome you to contact us and join us on this incredible journey.
Explore Our Research Areas
Speech Enhancement
Our expertise in speech enhancement technologies aims to improve communication in noisy environments, providing clearer, more accessible interactions for all.
Audio Analytics
We delve into the intricacies of audio analysis, focusing on tasks such as classifying background scenes and localizing acoustic events to advance understanding and applications in audio processing and interpretation.
Spatial Audio
We explore the immersive world of spatial audio, creating experiences that transcend traditional audio boundaries and enhance human perception focussing on realistic audio reproduction for group experiences in Virtual Reality.
Biomedical Signal Processing
We contribute novel signal processing approaches for pattern recognition and signal detection in sensor time-series. Our recent work encompasses cardio-vascular analysis based on signals obtained from laser doppler vibrometry (LDVs) and barcode detection using nanopores.
NEWS
Paper accepted at EUSIPCO 2025:
"Study of Lightweight Transformer Architectures for Single-Channel Speech Enhancement"
- Introduces LCT-GAN, a lightweight, causal transformer-based architecture for speech enhancement.
- Uses a Frequency-Time-Frequency (FTF) stacked transformer design to model global dependencies efficiently.
- Incorporates adversarial training to boost performance without increasing inference complexity.
- Performance Highlights:
- Achieves state-of-the-art results among lightweight models.
- Uses only 6% of the parameters compared to DeepFilterNet2 with similar performance.
- Saves 9% in parameters and 10% in MACs compared to CCFNet+(Lite), while performing better.
- Outperforms more complex baseline models on standard test datasets.
Paper accepted at EUSIPCO 2025:
"Array Agnostic Multi-channel Speech Presence Probability Estimation"
- Introduces a novel array-agnostic method for multi-channel speech presence probability (MC-SPP) estimation.
- Adapts a previous neural architecture to support a variable number of microphones with permutation invariance.
- Employs channel-wise convolution and Transformer layers, followed by Mean Pooling for channel fusion.
- Inserts Transform-Average-Concatenate layers to integrate array-level and channel-wise features.
- Uses a modified MVDR beamformer for spatial filtering.
- Performance Highlights:
- Comparable to array-fixed models on known microphone geometries.
- Better generalization to unseen array configurations.
- Superior performance under microphone permutation with lower complexity (model size and MACs).