Prof. Sharon Gannot Develops Artificial Intelligence Methods for Audio Processing

Modern audio technologies, from smart devices and communication systems to hearing technologies, must operate in complex acoustic environments, where multiple sound sources, noise, reverberation, and movement make reliable audio signal processing difficult.

The research project of Prof. Sharon Gannot, which won him a grant from the Israel Science Foundation (ISF), will develop new AI-based methods to learn the underlying structure of the acoustic environment from microphone-captured signals. “Although the acoustic information captured by multiple microphones is complex and high-dimensional, in many cases it is characterized by a hidden, simpler structure, determined by physical characteristics such as the positions of the sound sources and the microphones and the properties of the acoustic environment,” explains Prof. Gannot.

To achieve this, Prof. Gannot proposes developing manifold learning methods based on graph neural networks (GNNs) - in particular, graph convolutional networks (GCNs). “The goal is to uncover a simple, low-dimensional structure hidden within the complex, high-dimensional data, and in doing so to learn the relationships and internal structure of the data and to produce compact representations of the acoustic environment, while preserving clear physical meaning,” he clarifies.

Combining learned representations with methods from the world of signal processing is expected to enable the development of more accurate and robust algorithms for audio technologies, especially under complex and changing conditions. “The methods will be tested in key applications such as localization and tracking of sound sources, separation of simultaneously active speakers and sounds, beamforming, and acoustic echo cancellation,” says Prof. Gannot. “Beyond these applications, the project seeks to create a general framework that combines modern artificial intelligence with the physical principles of sound propagation, thereby advancing smart audio systems capable of operating more efficiently and reliably in realistic acoustic environments.”

This is the second consecutive year in which Prof. Gannot has won a grant from the Israel Science Foundation. Last year he won a joint grant from the ISF and the German Research Foundation (DFG). That grant, which is still active, focuses on developing deep-learning-based algorithms (including generative models) for smart hearing aids.