Conformer-enhanced hybrid neural architectures for sound event detection in driving scenarios

Engineering notes

Experimental results show that the proposed Conformer-integrated hybrid model outperforms the directly comparable CRNN-based baselines considered on this benchmark, achieving a Polyphonic Sound Detection Score (PSDS) of 47.3 and an F1 score of 65.7, demonstrating improved detection of complex acoustic events within the evaluated benchmark.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

Abstract The increasing adoption of smart vehicles demands safety systems capable of detecting critical acoustic events–such as emergency sirens, tires skidding, and car crashes–in dynamic driving environments. This paper studies a task-specific architectural refinement of a hybrid framework for sound event detection in driving scenarios by replacing the recurrent neck with compact Conformer blocks while retaining CNN, BEATs, and ATST-Frame feature pathways. The motivation is to improve temporal modeling by combining self-attention for broader context with convolution for local acoustic detail. Building on prior work using Convolutional Recurrent Neural Networks (CRNNs) and pre-trained models such as BEATs and ATST-Frame, we introduce Conformer blocks to better capture both global and local acoustic patterns, addressing challenges such as overlapping sounds and noisy conditions. Our approach leverages a dataset of 19,000 audio clips tailored for driving scenarios, augmented with techniques such as mixup and SpecAugment to improve training stability. The architecture fuses features from three parallel pathways through concatenation and projection, with Conformer blocks processing the unified representation to capture multi-scale acoustic characteristics. Experimental results show that the proposed Conformer-integrated hybrid model outperforms the directly comparable CRNN-based baselines considered on this benchmark, achieving a Polyphonic Sound Detection Score (PSDS) of 47.3 and an F1 score of 65.7, demonstrating improved detection of complex acoustic events within the evaluated benchmark. These results support the potential of Conformer-based architectures for sound event detection in smart-vehicle settings.

Links and sources

• PDF from original source

Comments