Scholarlink

Background <p>Environmental noise is a continuous challenge in audio anomaly detection, yet this study turns it into a chance to enhance classification systems’ resilience.</p> Purpose <p>This study aims at blending convolutional and Transformer-based backbones with temporal refinement, fostering noise-aware architectures that generalize across varied soundscapes.</p> Methods <p>We explore dl techniques that integrate typical feature representations (Mel Spectrogram, mfcc, and chromastft) with EfficientNet and swt backbones, enhanced by recurrent layers (gru and lstm) for temporal dependencies. The framework is tested on benchmark datasets from industrial (mimii), environmental (esc50), and wildlife (fsc22) domains.</p> Results <p>Results show strong performance with 98.86 ± 0.20% accuracy on mimii, 72.30 ± 2.41% on esc50, and 82.47 ± 0.98% on fsc22, validated statistically across folds. These findings highlight the hybrid cnn/Transformer-RNN models’ effectiveness in structured industrial settings, though diverse natural soundscapes remain tough. Analysis shows chromastft generally performs worse than melspectrogram and mfcc, stressing the importance of feature alignment with the acoustic domain.</p> Conclusions <p>The research achieves advanced performance in industrial sound classification, offers insights into feature selection and temporal modeling, and suggests future work involving geographically contextual audio for real-world industrial monitoring, urban planning, and environmental conservation.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

RNN-enriched Deep Transfer Learning Model for Audio-based Classification in Various Environmental Settings

  • 期刊论文

  • 发表日期：

    2026年1月21日

查看全文
  • Ahmad Qurthobi,
  • Robertas Damaševičius,
  • Sarmad Maqsood,
  • Rytis Maskeliūnas

摘要

Background

Environmental noise is a continuous challenge in audio anomaly detection, yet this study turns it into a chance to enhance classification systems’ resilience.

Purpose

This study aims at blending convolutional and Transformer-based backbones with temporal refinement, fostering noise-aware architectures that generalize across varied soundscapes.

Methods

We explore dl techniques that integrate typical feature representations (Mel Spectrogram, mfcc, and chromastft) with EfficientNet and swt backbones, enhanced by recurrent layers (gru and lstm) for temporal dependencies. The framework is tested on benchmark datasets from industrial (mimii), environmental (esc50), and wildlife (fsc22) domains.

Results

Results show strong performance with 98.86 ± 0.20% accuracy on mimii, 72.30 ± 2.41% on esc50, and 82.47 ± 0.98% on fsc22, validated statistically across folds. These findings highlight the hybrid cnn/Transformer-RNN models’ effectiveness in structured industrial settings, though diverse natural soundscapes remain tough. Analysis shows chromastft generally performs worse than melspectrogram and mfcc, stressing the importance of feature alignment with the acoustic domain.

Conclusions

The research achieves advanced performance in industrial sound classification, offers insights into feature selection and temporal modeling, and suggests future work involving geographically contextual audio for real-world industrial monitoring, urban planning, and environmental conservation.