Abstract
We study few-shot acoustic event detection (AED) in this paper. Few-shot learning enables detection of new events with very limited labeled data. Compared to other research areas like computer vision, few-shot learning for audio recognition has been under-studied. We formulate few-shot AED problem and explore different ways of utilizing traditional supervised methods for this setting as well as a variety of meta-learning approaches, which are conventionally used to solve few-shot classification problem. Compared to supervised baselines, meta-learning models achieve superior performance, thus showing its effectiveness on generalization to new audio events. Our analysis including impact of initialization and domain discrepancy further validate the advantage of meta-learning approaches in few-shot AED.
Introduction
Standard Acoustic Event Detection (AED) models are notoriously "data-hungry," typically requiring thousands of labeled clips to achieve high accuracy. This requirement presents a significant barrier for two main scenarios: detecting rare events (e.g., specific mechanical failures or rare wildlife) and personalizing systems for specific user environments (e.g., a unique doorbell sound). While few-shot learning has seen rapid advancement in computer vision and NLP, it remains under-explored in the audio domain. This paper formalizes the few-shot AED problem and evaluates whether meta-learning can outperform traditional transfer learning approaches.
Methods
The study compares two primary philosophies for solving the few-shot problem:
1. Supervised Baselines (Fine-Tuning)
- FT-All / FT-Linear: A model is pre-trained on a large set of known audio classes. Upon encountering a new task, either the entire model or just the final classification layer is fine-tuned using the limited available samples.
- Nearest Neighbor (NN): This approach uses a pre-trained feature extractor to map audio to a vector space. New sounds are classified based on their distance to the labeled "support" samples without additional training at test time.
2. Meta-Learning (Learning to Learn)
Instead of learning specific sounds, these models are trained on a variety of tasks to learn the general skill of distinguishing between any sounds.
- Prototypical Networks: These learn a metric space where classification is performed by computing distances to prototype representations. The authors found that averaging distances to support samples was more robust for audio than averaging the embeddings themselves.
- MetaOptNet: This replaces simple distance metrics with a differentiable linear SVM within the meta-learning loop.
- MAML (Model-Agnostic Meta-Learning): This optimizes initial parameters so the model can reach high performance on a new task after only a few gradient descent steps.
Experimental Setup
The researchers used a filtered subset of AudioSet, focusing on high-quality annotations. The final dataset consisted of 19,841 clips across 142 distinct event categories. These were split into 99 events for training, 21 for validation, and 21 for testing, with no overlap between the sets to ensure a true test of generalization.
Key Results
Meta-learning approaches consistently outperformed supervised baselines across both 1-shot and 5-shot scenarios. Prototypical Networks achieved the highest performance, reaching an AUC of 0.912 in the 5-shot setting.
In the ultra-low data 1-shot regime, the gap was even more pronounced. Prototypical Networks achieved an AUC of 0.833, while the standard fine-tuning approach (FT-All) struggled at 0.636 due to rapid overfitting. Interestingly, the Nearest Neighbor (NN) baseline outperformed fine-tuning, suggesting that a robust feature extractor is often more valuable than a complex classifier when data is scarce.
Analysis and Insights
Domain Sensitivity
A significant challenge in few-shot AED is domain mismatch. When models trained on one domain (e.g., household sounds) are tested on another (e.g., animal sounds), performance decreases across all methods. However, meta-learning methods maintained a clear lead over supervised baselines even in these cross-domain scenarios.
Distance-Based Robustness
The authors discovered that for audio, averaging distances to support samples was more effective than creating a single "mean" prototype. This is likely because "negative" samples in audio are extremely diverse, and a single average embedding fails to represent that variance effectively.
Conclusion
Meta-learning provides a superior framework for few-shot AED, enabling models to generalize to unseen audio categories far more effectively than traditional supervised fine-tuning. While domain sensitivity remains an open challenge, the results demonstrate that "learning to learn" is a promising path toward more flexible and data-efficient acoustic monitoring systems.
Summary updated