Speakers Localization Using Batch EM In Unfolding Neural Network

Researchers Rina Veler and Sharon Gannot have introduced a new AI architecture called a Batch-EM Unfolded Network, designed to solve a critical problem in audio processing: pinpointing where a speaker is located in a room. The core innovation is 'algorithm unfolding,' where they take the classical Expectation-Maximization (EM) statistical algorithm—used for finding maximum likelihood estimates—and embed its iterative steps directly into the layers of a neural network. This creates a hybrid model that is both interpretable, like traditional signal processing, and trainable end-to-end like a deep learning system. The result is a method that mitigates the notorious sensitivity of EM to its starting point, leading to faster and more reliable convergence.

In practical tests, the unfolded network demonstrated superior accuracy and robustness compared to the standalone Batch-EM algorithm, especially in difficult acoustic environments with strong reverberation (echoes). This advancement directly tackles a major hurdle for real-world audio AI, such as in smart speakers, video conferencing systems, and hearing aids, where reflections and noise often degrade performance. By providing a more reliable way to localize sound sources, this research, presented at ICSEE 2026, paves the way for voice interfaces that work flawlessly in kitchens, living rooms, and conference halls, not just in acoustically treated labs.