New Deep Learning Model Improves Sound Event Localization and Classification Across Multiple Microphone Arrays

Global: New Deep Learning Model Improves Sound Event Localization and Classification Across Multiple Microphone Arrays

Researchers have introduced a deep‑learning approach that simultaneously localizes and classifies sound events using multiple microphone arrays, addressing signal attenuation and environmental noise that limit single‑array systems. The study, posted on arXiv in March 2024, reports performance gains over existing techniques in both classification accuracy and localization precision.

Background and Motivation

Sound event localization and classification (SELC) is a growing focus within wireless acoustic sensor networks, yet most current implementations rely on a single microphone array. Such configurations are vulnerable to reduced monitoring range and heightened susceptibility to ambient noise, while multi‑array solutions have traditionally emphasized only source localization, leaving classification underexplored.

Method Overview

The proposed method introduces a novel “Soundmap” feature that captures spatial information across multiple frequency bands, complemented by acoustic representations derived from a Gammatone filter tuned for outdoor environments. Attention mechanisms are integrated to learn channel‑wise relationships and temporal dependencies within these features, enabling the model to produce both location coordinates and class labels.

Experimental Evaluation

Simulated datasets were generated with varying noise levels, monitoring area sizes, array geometries, and source positions to assess robustness. Comparative experiments demonstrate that the new approach outperforms state‑of‑the‑art baselines on both classification and localization metrics, indicating a measurable advantage in diverse acoustic scenarios.

Error Analysis

Further analysis attributes residual errors primarily to reverberation effects and overlapping sound sources, suggesting that future refinements in feature extraction and model architecture could mitigate these challenges.

Potential Applications

By extending the effective monitoring range and improving classification reliability, the technique holds promise for applications such as urban sound surveillance, wildlife monitoring, and industrial safety systems, pending validation on real‑world acoustic recordings.

This report is based on information from arXiv, licensed under Academic Preprint / Open Access. Based on the abstract of the research paper. Full text available via ArXiv.