Abstract
This paper presents a novel dual-input Convolutional Neural Network (CNN) for Sound Event Detection (SED) in real-world environments. By processing both short-term local features and long-term global context, the model effectively disambiguates overlapping sounds and varying acoustic backgrounds. Coupled with optimization techniques such as class-wise early stopping and adaptive thresholding, the proposed approach significantly outperforms the DCASE 2017 baseline, achieving a 67.0% F-score.
Introduction
Sound Event Detection (SED) involves identifying the type and temporal boundaries of acoustic events within an audio stream. Real-world audio presents significant challenges, including overlapping sounds, background noise, and severe class imbalance. This study addresses Task 3 of the DCASE 2017 Challenge, focusing on detecting six diverse sound classes—ranging from "people walking" to "large vehicles"—within a limited dataset of 24 recordings.
Methodology
The core innovation is a "Multiple-Input CNN" architecture designed to capture audio features at two distinct temporal scales:
- Short-term Input: A ~4-second stereo window is converted into a log-mel spectrogram to capture local acoustic characteristics.
- Long-term Input: The average log-mel spectrogram of the entire 3–5 minute recording provides global context, helping the model understand the broader acoustic environment.
These inputs are processed through separate convolutional paths and merged via addition before passing through fully connected layers for final classification. To enhance training, the authors employed frequent validation (every 20 mini-batches), class-wise early stopping to handle varying learning rates, and adaptive thresholds to optimize detection for imbalanced classes.
Experimental Setup
The model was evaluated using the TUT sound events 2017 dataset, which consists of 24 stereo recordings (3–5 minutes each) at 44.1kHz. The six target classes include: brakes squeaking, car, children, large vehicle, people speaking, and people walking. Performance was measured using the standard 4-fold cross-validation provided by the DCASE organizers.
Results and Discussion
The proposed model demonstrated substantial improvements over the DCASE 2017 baseline. The overall Error Rate (ER) decreased from 0.69 to 0.51, while the F-score improved from 56.7% to 67.0%.
Analysis of class-wise performance reveals that common sounds like "car" and "large vehicle" were detected with high accuracy (F-scores in the 60-70s). In contrast, sparse events like "brakes squeaking" and "children" remained challenging, highlighting the persistent difficulty of extreme data imbalance. The inclusion of global context (logAvgMel) was identified as a critical factor in improving the model's ability to distinguish between similar-sounding events in different environments.
Conclusion
The Multiple-Input CNN effectively leverages global context and specialized training strategies to advance state-of-the-art sound event detection. While the model significantly outperformed the baseline, limitations remain regarding training volatility and the potential loss of information during spectrogram conversion. Future work will explore more robust data sampling techniques to address class imbalance and investigate end-to-end learning architectures that process raw waveforms directly.
Summary updated