Sub-Band Spectral Matching with Localized Score Aggregation for Robust Anomalous Sound Detection

Detecting subtle deviations in noisy acoustic environments is central to anomalous sound detection (ASD). A common training-free ASD pipeline temporally pools frame-level representations into a band-preserving feature vector and scores anomalies using a single nearest-neighbor match. However, this global matching can inflate normal-score variance through two effects. First, when normal sounds exhibit band-wise variability, a single global neighbor forces all bands to share the same reference, increasing band-level mismatch. Second, cosine-based matching is energy-coupled, allowing a few high-energy bands to dominate score computation under normal energy fluctuations and further increase variance. We propose BEAM, which stores temporally pooled sub-band vectors in a memory bank, retrieves neighbors per sub-band, and uniformly aggregates scores to reduce normal-score variability and improve discriminability. We further introduce a parameter-free adaptive fusion to better handle diverse temporal dynamics in sub-band responses. Experiments on multiple DCASE Task 2 benchmarks show strong performance without task-specific training, robustness to noise and domain shifts, and complementary gains when combined with encoder fine-tuning.

💡 Research Summary

The paper addresses a fundamental weakness in many training‑free anomalous sound detection (ASD) pipelines: the use of a single global nearest‑neighbor (NN) match on a clip‑level embedding. While this design is compact and works well with large pre‑trained audio encoders, it inflates the variance of normal‑sound scores under domain shift or noisy conditions. The authors identify two specific mechanisms behind this variance inflation. First, when normal sounds exhibit band‑wise variability (e.g., some frequency bands change more than others due to background noise), a single global NN forces all bands to share the same reference, causing a mismatch for many bands. Second, cosine similarity is energy‑coupled; high‑energy bands dominate both the selection of the global reference and the final score, so fluctuations in those bands translate directly into larger score variance.

To mitigate these issues, the authors propose BEAM (Band‑wise Equalized Anomaly Measure). BEAM partitions each clip embedding into a set of contiguous sub‑bands (defined by a window size and stride) and stores the sub‑band vectors of all normal training clips in a band‑aligned memory bank. At inference, each sub‑band of a test clip queries only its own band‑specific memory, retrieving the nearest neighbor using cosine distance. The per‑band distances are then uniformly averaged to produce the anomaly score. This uniform aggregation removes the energy‑coupling effect, while the per‑band retrieval eliminates the tied‑reference problem. Because sub‑band distances can have different scales across bands and domains, the authors further apply a local density normalization (LDN): each distance is divided by the average distance of the K nearest neighbors around the matched reference within the same band, providing a band‑specific scale calibration before aggregation.

AdaBEAM extends BEAM with Dynamic Mean‑Max (DMM) integration to handle diverse temporal dynamics. Two parallel clip‑level representations are built: one from temporal mean pooling (capturing stable spectral structure) and another from temporal max pooling (emphasizing transient high‑energy peaks). Both are processed through the BEAM pipeline, yielding two scores that are fused by a simple parameter‑free rule (by default the arithmetic mean). This fusion captures both steady‑state and transient information, improving robustness in noisy or domain‑shifted scenarios.

Theoretical analysis shows that the global cosine score can be decomposed into band‑wise contributions weighted by energy. The authors prove that uniform aggregation of independent band scores reduces the variance of the normal‑score distribution. Using signal detection theory, they derive a sufficient condition under which the reduction in variance outweighs any potential loss in mean separation, guaranteeing an increase in detection sensitivity (d′).

Experiments are conducted on multiple DCASE Task 2 benchmarks covering several machine types and noise levels. The authors evaluate three families of front‑ends: handcrafted Log‑Mel, MFCC, and LPC spectra; deep spectrogram embeddings from large‑scale pre‑trained encoders such as BEATs; and fine‑tuned versions of these encoders. In the training‑free setting, BEAM and AdaBEAM consistently outperform global‑matching baselines, achieving 3–5 percentage‑point gains in AUC/p‑AUC, especially under low SNR conditions where global methods suffer from inflated score variance. AdaBEAM’s DMM fusion further improves performance by effectively combining stable and transient cues. When the encoder is fine‑tuned on target‑machine data, BEAM still provides additional 1–2 percentage‑point improvements over the corresponding global‑matching fine‑tuned system, demonstrating that the proposed sub‑band strategy complements supervised adaptation.

In summary, the paper presents a clear diagnosis of why global NN matching is suboptimal for ASD, introduces a simple yet powerful sub‑band memory and uniform aggregation framework (BEAM) together with a parameter‑free dynamic fusion (AdaBEAM), provides theoretical justification for variance reduction, and validates the approach across diverse features and realistic acoustic conditions. The methods require only storage of sub‑band descriptors and no additional training, making them attractive for real‑time industrial monitoring where rapid deployment and robustness to environmental changes are essential.