Temporal Pooling Strategies for Training-Free Anomalous Sound Detection with Self-Supervised Audio Embeddings

Kevin Wilkinghoff, Sarthak Yadav, Zheng-Hua Tan

Abstract

Training-free anomalous sound detection (ASD) using pre-trained audio embedding models has gained traction due to its robustness under domain shifts and minimal data requirements. However, the standard reliance on temporal mean pooling often overlooks critical localized acoustic deviations. This paper introduces Relative Deviation Pooling (RDP) and a hybrid RDP+GeM strategy to better capture these anomalies. Our systematic evaluation across multiple state-of-the-art models demonstrates that these strategies consistently outperform mean pooling, achieving state-of-the-art results on benchmark datasets, including the DCASE2025 ASD challenge.

Introduction

Anomalous Sound Detection (ASD) is vital for industrial condition monitoring, identifying machine failures from acoustic signals. Recent "training-free" approaches leverage large-scale self-supervised models (e.g., BEATs, OpenL3, EAT) to extract embeddings. A critical, yet often overlooked, step in this pipeline is temporal pooling: aggregating a sequence of embeddings into a single representative vector. While mean pooling is the default, it tends to smooth out short-lived, informative anomalies. This research investigates alternative pooling strategies to enhance the sensitivity of training-free ASD systems.

Proposed Pooling Strategies

We propose two adaptive pooling methods designed to emphasize unusual temporal events:

  1. Relative Deviation Pooling (RDP): RDP assigns weights to temporal frames based on their deviation from the global temporal mean. By emphasizing frames that are "different" from the average, RDP focuses on acoustic events that are likely to represent anomalies.
  2. Hybrid RDP + GeM: This strategy combines RDP weights with Generalized Mean (GeM) pooling. GeM provides a tunable aggregation that spans the range from mean to max pooling. The combination leverages both the adaptive weighting of RDP and the robust feature selection of GeM.
Performance change of the proposed RDP+GeM strategy across different embedding models on the development set as a function of the hyperparameter γ.

Experimental Evaluation

The strategies were evaluated using the DCASE ASD benchmark datasets (2020–2025), which emphasize domain generalization. We tested four state-of-the-art embedding models: OpenL3, BEATs, EAT, and Dasheng. Notably, for the EAT model, we identified that hard thresholding and spike suppression are crucial pre-processing steps to optimize anomaly detection performance.

Key Findings

  • Pooling as a Performance Lever: In training-free pipelines, temporal pooling is a primary driver of performance. Optimized pooling can yield improvements comparable to upgrading the underlying embedding model.
  • Model Sensitivity: The benefit of RDP varies by model. Models like BEATs, which retain high temporal heterogeneity, show significant gains, whereas more stable models like EAT benefit less.
  • Superiority over Mean Pooling: RDP and RDP+GeM consistently surpassed standard mean pooling across all tested models and datasets.
Comprehensive comparison of pooling strategies (GWRP, RDP, GeM, RDP+GeM) across development and evaluation sets, showing performance sensitivity to hyperparameters.

Results

The proposed RDP+GeM strategy achieved state-of-the-art performance for training-free ASD. Notably, on the DCASE2025 dataset, a system using BEATs with RDP+GeM reached a harmonic mean AUC of 63.2%, outperforming all previously reported supervised systems and ensembles. This result challenges the assumption that training-free methods are inherently inferior to trained ones, suggesting that suboptimal aggregation was a key limiting factor.

Conclusion

Temporal pooling is a critical component of training-free ASD. By introducing RDP and RDP+GeM, we provide effective tools for capturing localized anomalies in audio embeddings. These methods offer a simple yet powerful way to boost the performance of ASD systems without requiring additional training data or complex architectures.

Summary updated

Abstract

Training-free anomalous sound detection (ASD) based on pre-trained audio embedding models has recently garnered significant attention, as it enables the detection of anomalous sounds using only normal reference data while offering improved robustness under domain shifts. However, existing embedding-based approaches almost exclusively rely on temporal mean pooling, while alternative pooling strategies have so far only been explored for spectrogram-based representations. Consequently, the role of temporal pooling in training-free ASD with pre-trained embeddings remains insufficiently understood. In this paper, we present a systematic evaluation of temporal pooling strategies across multiple state-of-the-art audio embedding models. We propose relative deviation pooling (RDP), an adaptive pooling method that emphasizes informative temporal deviations, and introduce a hybrid pooling strategy that combines RDP with generalized mean pooling. Experiments on five benchmark datasets demonstrate that the proposed methods consistently outperform mean pooling and achieve state-of-the-art performance for training-free ASD, including results that surpass all previously reported trained systems and ensembles on the DCASE2025 ASD dataset.
BibTeX
@misc{wilkinghoff2026temporal,
  author = {Kevin Wilkinghoff and Sarthak Yadav and Zheng-Hua Tan},
  title = {{Temporal Pooling Strategies for Training-Free Anomalous Sound Detection with Self-Supervised Audio Embeddings}},
  year = {2026},
  eprint = {2603.04605},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2603.04605},
}
Figure