[Paper Review] Cross-task learning for audio tagging, sound event detection spatial localization: DCASE 2019 baseline systems

Qiuqiang Kong, Yin Cao+4 more|arXiv (Cornell University)|Apr 11, 2019

Music and Audio Processing4 citations

TL;DR

This paper proposes generic cross-task baseline systems using 5-, 9-, and 13-layer convolutional neural networks (CNNs) for audio tagging, sound event detection, and spatial localization in the DCASE 2019 challenge. The 9-layer CNN with average pooling achieves the best performance across most tasks, demonstrating that model depth and pooling type significantly influence cross-task generalization without task-specific tuning.

ABSTRACT

The Detection and Classification of Acoustic Scenes and Events (DCASE) 2019 challenge focuses on audio tagging, sound event detection and spatial localisation. DCASE 2019 consists of five tasks: 1) acoustic scene classification, 2) audio tagging with noisy labels and minimal supervision, 3) sound event localisation and detection, 4) sound event detection in domestic environments, and 5) urban sound tagging. In this paper, we propose generic cross-task baseline systems based on convolutional neural networks (CNNs). The motivation is to investigate the performance of a variety of models across several tasks without exploiting the specific characteristics of the tasks. We looked at CNNs with 5, 9, and 13 layers, and found that the optimal architecture is task-dependent. For the systems we considered, we found that the 9-layer CNN with average pooling is a good model for a majority of the DCASE 2019 tasks.

Motivation & Objective

• To develop generic, reusable baseline models for multiple audio understanding tasks without exploiting task-specific characteristics.
• To investigate how model architecture depth (5, 9, 13 layers) affects performance across diverse audio tasks.
• To evaluate the impact of pooling strategy (specifically average pooling) on cross-task generalization in audio representation learning.
• To establish a robust, transferable baseline for future research in audio tagging, sound event detection, and spatial localization.
• To provide insights into optimal model design for multi-task audio understanding in low-supervision and noisy-label settings.

Proposed method

• Designing three CNN architectures with 5, 9, and 13 layers to evaluate depth effects on cross-task performance.
• Using average pooling in the 9-layer CNN as a key architectural component to improve generalization across tasks.
• Training the same model architecture across five distinct DCASE 2019 tasks: acoustic scene classification, audio tagging with noisy labels, sound event localization and detection, detection in domestic environments, and urban sound tagging.
• Applying a shared training and inference pipeline across tasks to ensure fair comparison and avoid task-specific adaptation.
• Evaluating model performance using standard DCASE 2019 metrics, including F1-score and localization accuracy.
• Selecting the 9-layer CNN with average pooling as the optimal baseline after empirical comparison across tasks.

Experimental results

Research questions

• RQ1 How does model depth (5, 9, 13 layers) influence performance across diverse audio tagging and detection tasks?
• RQ2 Does average pooling improve cross-task generalization compared to other pooling strategies in audio representation learning?
• RQ3 Can a single, generic CNN architecture achieve strong performance across multiple DCASE 2019 tasks without task-specific adaptation?
• RQ4 Which CNN architecture provides the best trade-off between complexity and performance for multi-task audio understanding?
• RQ5 How does the 9-layer CNN with average pooling compare to deeper or shallower models in terms of robustness and generalization across noisy and minimal-supervision settings?

Key findings

• The 9-layer CNN with average pooling achieved the best overall performance across the majority of DCASE 2019 tasks.
• Model performance was found to be task-dependent, indicating that optimal architecture varies with task characteristics.
• The 9-layer CNN outperformed both the shallower 5-layer and deeper 13-layer models in terms of generalization and robustness.
• Average pooling contributed to improved performance, particularly in tasks requiring precise localization and detection.
• The proposed generic baseline system demonstrated strong transferability across tasks with minimal architectural modification.
• The study confirms that architectural choices such as depth and pooling type are critical for cross-task learning in audio understanding.