Abstract
Audio Event Detection (AED) involves identifying specific sound events within continuous, noisy audio streams. Traditional approaches often struggle with data imbalance and the neglect of temporal structures. This paper proposes a robust AED system that leverages specialized loss functions rather than just architectural modifications. By employing a weighted loss to mitigate background noise dominance and a multi-task loss to model event boundaries, the system achieved an 88.3% F1-score in the DCASE 2017 challenge, a substantial improvement over the 64.1% baseline.
Introduction
The primary goal of AED is to determine the "what" and "when" of sounds like baby cries or gunshots. Current research faces three main challenges: an over-reliance on architectures borrowed from other domains, severe data imbalance where background noise outweighs rare events, and the failure to account for the temporal onset and offset of sounds. This work addresses these by focusing on the mathematical formulation of the learning objectives.
Methodology
The proposed system follows a two-stage pipeline consisting of background rejection and event classification.
Phase-Aware Signal Enhancement
Audio signals are first cleaned using an enhancement technique that considers both magnitude and phase. By estimating the clean phase of harmonic sounds, the system better reconstructs signals from background noise compared to magnitude-only methods.
Weighted Loss for Background Rejection
To handle the "needle in a haystack" problem of rare events, a binary classifier is trained with a weighted cross-entropy loss. False negatives are penalized ten times more heavily than false positives, ensuring the model remains sensitive to rare occurrences amidst dominant background noise.
Multi-task Loss for Event Classification
For the classification stage, the network is trained on three concurrent tasks:
- Classification: Identifying the sound category.
- Distance Estimation: Predicting the temporal distance from the current frame to the event's onset and offset.
- Confidence: A term based on the Intersection over Union (IoU) of predicted versus ground truth boundaries.
This multi-task approach acts as an implicit regularizer, forcing the network to learn the "shape" of the sound event.
Experimental Results
The system was evaluated on the DCASE 2017 Challenge Task 2 dataset, featuring rare events (baby cry, glass break, gun shot) mixed with TUT Acoustic Scenes 2016 background noise. The proposed approach significantly outperformed the DCASE baseline:
- F1-Score: Improved from 64.1% to 88.3% on the evaluation set.
- Error Rate: Reduced from 0.64 to 0.22.
- Challenge Ranking: The system ranked 3rd out of 13 participating teams.
Discussion: CNNs vs. DNNs
The study revealed distinct advantages for different architectures. CNNs excelled at detecting baby cries, likely due to their ability to model vocal-tract variations. In contrast, DNNs performed better on glass breaks and gunshots, where preserving high-frequency details without the "blurring" effect of convolutional pooling proved beneficial.
Conclusion and Future Work
By shifting focus to loss function design, this work demonstrates that significant gains in AED performance are possible without complex architectural changes. Future work will explore using temporal distance estimates for real-time early event detection and extending the system to handle multi-event environments.
Summary updated