Related Experiment Video
Updated: Oct 8, 2025
08:25
Combining Eye-tracking Data with an Analysis of Video Content from Free-viewing a Video of a Walk in an Urban Park Environment
Published on: May 7, 2019
9.2K
Sound Event Detection by Pseudo-Labeling in Weakly Labeled Dataset
Chungho Park1, Donghyeon Kim1, Hanseok Ko1
1Department of Electronics and Electrical Engineering, Korea University Seoul, Seoul 136-713, Korea.
Sensors (Basel, Switzerland)
|December 28, 2021
Summary
This study introduces an improved weakly labeled sound event detection model using gated linear units and dilated convolutions. The novel approach enhances accuracy in real-world audio by better handling unknown sounds and achieving significant performance gains.
Area of Science:
- Audio signal processing
- Machine learning for acoustics
- Deep learning for sound event detection
Background:
- Weakly labeled sound event detection (WSED) is crucial for efficient dataset creation.
- Current deep learning models struggle with real-time audio due to limitations in feature extraction and handling unknown sounds.
- Convolutional Neural Networks (CNNs) in WSED often lack sufficient receptive fields and fail to prioritize important features.
Purpose of the Study:
- To develop a more robust and efficient WSED model for real-world audio streams.
- To address limitations in feature importance highlighting and receptive field size in existing CNN-based WSED approaches.
- To improve the model's ability to distinguish between target and unknown sound events.
Main Methods:
- Implemented a novel model incorporating Gated Linear Units (GLU) and dilated convolutions to enhance feature representation.
- Introduced pseudo-label-based learning with 'noise labels' and 'noise loss' to effectively classify and separate unknown sound content.
- Utilized a combination of DCASE 2018 task 1 (acoustic scenes) and task 2 (sound events) data for experimentation.
Main Results:
- The proposed sound event detection (SED) model achieved state-of-the-art F1 scores: 59.7% at 0 SNR, 64.5% at 10 SNR, and 65.9% at 20 SNR.
- Demonstrated significant performance improvements over the baseline model, with increases of 17.7%, 16.9%, and 16.5% at the respective SNR levels.
- The pseudo-labeling strategy effectively reduced the impact of unknown environmental sounds on detection accuracy.
Conclusions:
- The proposed WSED model effectively overcomes limitations of previous methods by improving feature learning and handling of unknown audio content.
- The integration of GLU, dilated convolutions, and pseudo-labeling offers a promising direction for robust sound event detection in challenging acoustic environments.
- The experimental results validate the model's superior performance and its potential for practical applications in real-time audio analysis.
