The approach was developed by Zahra Abolfazli and Hamid Reza Abutalebi and builds on the DCASE baseline architecture. Its central design idea is borrowed from human hearing: people typically identify a sound event before resolving its detailed spatial position. Following that principle, the new framework feeds information from the sound event detection branch into the localization branch, so that knowing what the sound is helps the network pin down where it is.

The mechanism that carries this information across is an Attentional Feature Fusion block, which combines relevant audio representations to strengthen spatial and temporal feature extraction. The model is trained with a dual-component loss function designed to improve temporal feature learning and enhance spatial localization performance at the same time.

Experimental results reported in the paper show that the framework offers two operating points. A balanced configuration, which uses recurrent modeling in feature extraction, outperforms the baseline across all metrics. A second, localization-oriented configuration, which places recurrent modeling in the detection branch, was selected for detailed evaluation because the authors prioritized spatial localization and distance estimation as the primary objective of the work.

The localization-oriented configuration delivers the most striking gains. It reduces angular error from 36.9 degrees to 25.7 degrees and relative distance error from 0.33 to 0.267 compared with the baseline. These are substantial improvements in the paper's two spatial measures: direction-of-arrival error and relative distance error.

There is a trade-off. Detection performance, measured by F-score, drops slightly, from 13.0 percent for the baseline to 11.7 percent for the new configuration, a change the authors describe as broadly comparable. The model also grows from approximately 700,000 parameters to about 1.07 million, a modest increase by modern standards but not a negligible one.

Even so, the authors argue the enlarged model remains substantially more compact than many high-complexity state-of-the-art systems, yielding what they characterize as a favorable trade-off between model complexity and spatial estimation performance. In other words, the system buys much better localization accuracy at a small cost in detection score and parameter count.

The practical significance lies in where such systems might run.

The human auditory system’s habit of recognizing a sound before localizing it proved to be a useful inductive bias, and the attention-based fusion mechanism shows how detection information can be routed to sharpen spatial estimates rather than treated as a separate output.

As with any architecture study, the reported figures come from the authors’ own experimental comparisons against the DCASE baseline within the SELDD task, and the localization-oriented configuration was chosen specifically to prioritize spatial accuracy. Within that scope, the results indicate that a human-inspired ordering of detection before localization, implemented through attentional feature fusion and a dual-component loss, can meaningfully improve direction-of-arrival and distance estimation without demanding a large model.

Subject of Research: A human-inspired neural network for sound event localization, detection and distance estimation

Article Title: A human-inspired SELDD approach for improved direction-of-arrival and distance estimation

Article References: Abolfazli, Z., & Abutalebi, H. R. (2026). A human-inspired SELDD approach for improved direction-of-arrival and distance estimation. Multimedia Tools and Applications, 85(11), Article 825. https://doi.org/10.1007/s11042-026-21984-2

Image Credits: AI Generated

DOI: 10.1007/s11042-026-21984-2

Keywords: sound event localization, SELDD, direction-of-arrival, distance estimation, DCASE baseline, attentional feature fusion, neural network, audio sensing, machine learning, spatial audio, model compactness, human auditory system