Semi-Supervised GNN for Sound Source Localization with Prediction Intervals

This paper presents a novel semi-supervised Graph neural network (GNN) framework for Sound source localization (SSL) in noisy and reverberant environments, complemented by a statistical method for constructing prediction intervals (PIs) that capture the true source position with high probability. This combination addresses two key limitations of existing SSL approaches: the reliance on large labeled datasets for training and the lack of reliability that limit their practical usage in real-world applications. Our approach builds on the assumption that acoustic features in a given environment lie on an underlying manifold, where nearby samples correspond to similar positions. We discretize this manifold into a graph, leveraging a small set of labeled samples together with a larger pool of unlabeled data. An integrated autoencoder further learns compact and informative representations from both labeled and unlabeled samples. To ensure reliability, we adopt the Jackknife+ procedure, which provides valid PIs without requiring extra calibration data and with manageable computational cost. We demonstrate the effectiveness of our approach through improved localization performance over baselines, along with its ability to generate tight and reliable PIs.