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

Phurich Saengthong; Takahiro Shinozaki

cs.SDMar 14, 2026

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

Authors: Phurich Saengthong, Takahiro Shinozaki

Organizations: Department of Information and Communications Engineering, Institute of Science Tokyo, Japan

Abstract

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.

Figures & tables

Fig. 1: Overview of BEAM and AdaBEAM . Normal audio is mapped to handcrafted or deep features, sliced into sub-bands, and stored in a shared memory bank. At test time, each query sub-band is matched within its band-aligned memory, and the resulting band scores are uniformly aggregated to produce the anomaly score. AdaBEAM adds Dynamic Mean–Max (DMM) fusion by building mean- and max-pooled sub-band memories, scoring both views, and combining the resulting clip scores with a simple parameter-free rule.

Dataset	Split	Source domain		Target domain
(sec-dev & sec-eval)		Normal	Anomalous	Normal	Anomalous
DCASE2020T2 [ 5 ]	Train	-	-	$\leq 1000$	-
	Test	-	-	$\leq 400$	$\leq 200$
DCASE2023T2 [ 15 ]	Train	990	-	10	-
	Test	50	50	50	50
DCASE2024T2 [ 16 ]	Train	990	-	10	-

TABLE I: Per-section statistics of each machine type in the considered ASD datasets.

Feature	Dataset	$(C,s)$	$N_{b}$
Log-Mel	DCASE2020T2	(38, 38)	4
	DCASE2023T2	(38, 38)	4
	DCASE2024T2	(76, 76)	2
MFCC	All	(20, 20)	5
LPC spectrum	All	(3200, 3200)	3
BEATs [ 19 ]	All	(768, 768)	8

TABLE II: BEAM windowing parameters (

C

: window dim,

s

: stride,

N_{b}

: number of windows (bands)).

	DCASE2020T2		DCASE2023T2 (DG)		DCASE2024T2 (DG)
Method	sec-dev (Amean)	sec-eval (Amean)	sec-dev (Hmean)	sec-eval (Hmean)	sec-dev (Hmean)	sec-eval (Hmean)
Systems requiring task-specific training
AE [ 5 , 15 , 16 ]	66.6	70.0	56.9	61.1	55.4	56.5
AE-IDNN [ 29 ]	71.5	-	-	-	-	-
ANP-IDNN [ 30 ]	71.8	75.6	-	-	-	-
PAE [ 31 ]	74.2	-	-	-	-	-

TABLE III: Comparison of our training-free frontend systems with previously reported results that do not use meta-information (auxiliary labels) for training or fine-tuning. This setting assumes access only to normal data from the same machine type (in both source and target domains); thus task-specific training on normal data (e.g., AE-style methods) is permitted, while methods that rely on auxiliary labels are excluded. Official scores are reported on DCASE2020T2, DCASE2023T2, and DCASE2024T2; we use the arithmetic mean (Amean) for DCASE2020T2 and the harmonic mean (Hmean) for DCASE2023T2 and DCASE2024T2.

Fig. 2: Quantitative results for the theoretical analysis using BEATs iter3 features, averaged across machine types for each benchmark (DCASE2020T2, DCASE2023T2, and DCASE2024T2). (Left) Variance ratio (no LDN). (Center-left) Mean gap ratio (no LDN). (Center-right)

d^{\prime}

ratio (no LDN). (Right)

d^{\prime}

ratio with Local Density Normalization (LDN), including additional results with DMM.

Prior work with global matching design
	DCASE2020T2		DCASE2023T2 (DG)		DCASE2024T2 (DG)
Method	sec-dev (Amean)	sec-eval (Amean)	sec-dev (Hmean)	sec-eval (Hmean)	sec-dev (Hmean)	sec-eval (Hmean)
SSL4ASD w/ LDN [ 12 , 13 ]	90.7	90.2	68.4	68.0	62.0	54.7
+ 10 ensembles	94.2	93.3	71.3	72.4	65.2	56.5
AnoPatch [ 8 ] (BEATs Full FT)	90.9	94.3	64.2	74.2	62.5 $\pm$ 0.8	65.6 $\pm$ 1.1
+ pseudo labeling [ 54 ]	-	-	-	-	64.1 $\pm$ 0.6	66.0 $\pm$ 0.7

TABLE IV: Comparison of fine-tuned frontend systems that use meta-information (auxiliary labels) during task-specific training, together with previously reported results from the literature. Official scores are reported on DCASE2020T2, DCASE2023T2, and DCASE2024T2; scores for our approach are averaged over five independent trials. We use the arithmetic mean (Amean) for DCASE2020T2 and the harmonic mean (Hmean) for DCASE2023T2 and DCASE2024T2.

Fig. 3: Comparison of sub-band window sizes of BEAM on handcrafted and deep features (BEATs) in terms of AUC across DCASE2020T2, DCASE2023T2, and DCASE2024T2. Window size is reported as a fraction of the feature length along the windowed axis. For handcrafted features, we use non-overlapping windows for 10%–40% (stride equals window length); for 50%–90%, we use a two-window full-coverage rule with an end-aligned final window (equivalently, stride

=F-C

), which may overlap with the first window. For BEATs, windows are formed by grouping a fixed number of patch embeddings (8 patches total), thus only window sizes aligned to the patch grid are meaningful. Dotted horizontal lines show the baseline Tmean result.

Fig. 4: Effect of varying

K

in sub-band density normalization on Official Scores across DCASE2020T2, DCASE2023T2, and DCASE2024T2. Results are shown for Tmean, Tmax, and DMM variants using BEATs features.

Baseline w/o Local Density Normalization (apply [ 9 ] )
	DCASE2020T2		DCASE2023T2 (DG)		DCASE2024T2 (DG)
Method	sec-dev	sec-eval	sec-dev	sec-eval	sec-dev	sec-eval
Tmean	75.8	77.5	62.6	71.7	56.3	59.0
w/ Sub-Band
Average	77.5	80.3	63.1	72.7	59.4	60.0
Maximum	78.0	80.7	62.7	69.9	57.3	59.4

TABLE V: Comparison of different BEAM sub-band scoring approaches using BEATs features. Higher Official Scores indicate better performance.

Fig. 5: Comparison of features on the MIMII dataset.

Frontend	Output dim	Feat. (ms)	E2E (ms)	MACs/Params (G/M)
Log-Mel (Tmean)	128	$3.95\pm 0.21$	$6.20\pm 0.25$	N/A
MFCC (Tmean)	90	$3.95\pm 0.01$	$7.95\pm 1.08$	N/A
LPC spectrum	8000	$31.34\pm 0.05$	$41.11\pm 0.15$	N/A
BEATs iter3 (Tmean, CPU)	6144	$271.15\pm 6.06$	$494.02\pm 6.19$	$45.01/90.71$
BEATs iter3 (Tmean, GPU)	6144	$13.60\pm 0.14$	$17.91\pm 0.14$	$45.01/90.71$

TABLE VI: Computational cost per 10 s clip (mean

\pm

std over

n_{\mathrm{runs}}=5

). Feature-extraction time measures waveform

\rightarrow

f, and end-to-end (E2E) time measures waveform

\rightarrow

anomaly score, including

1

-NN scoring against a memory bank of

500

normal clips. Handcrafted frontends are computed on CPU only in this benchmark. CPU: AMD Ryzen 5 5600X. GPU: NVIDIA GeForce RTX 3090. For BEATs, the encoder complexity is reported as MACs/Params (G/M).

Appendix figures & tables1 asset

Supplementary material from the paper’s appendix.

Appendix

Dataset	Band size	$\mathrm{Var(Sub)}/\mathrm{Var(Uni)}$	$\mathrm{Var(Uni)}/\mathrm{Var(Glob)}$	$\mathrm{Var(Sub)}/\mathrm{Var(Glob)}$	$\Delta\mu_{\mathrm{Glob}}$	$\Delta\mu_{\mathrm{Sub}}$	$d^{\prime}_{\mathrm{Glob}}$	$d^{\prime}_{\mathrm{Sub}}$
BEATs iter3
DCASE2020	768	0.485	1.023	0.496	0.009	0.008	2.023	2.398
	1536	0.609	1.011	0.616	0.009	0.008	2.023	2.314
	2304	0.686	1.002	0.686	0.009	0.009	2.023	2.258
	3072	0.766	1.006	0.770	0.009	0.009	2.023	2.183
	6144 (global)	1.000	1.000	1.000	0.009	0.009	2.023	2.023

Appendix

TABLE VII: Variance-ratio measures, mean-gap terms, and sensitivity index (

d^{\prime}

) across band sizes for band-preserved features; all methods are compared without LDN.

Explore similar work

Scoring Backends Matter More Than Pooling: A Systematic Study of Training-Free Anomalous Sound Detection under Domain Shift
Jun 17, 2026Jingwen Zhou, Mingzhe WangDistribution Shift Unsupervised Anomaly Detection
Xidian University, Xi’an, China
Temporal Pooling Strategies for Training-Free Anomalous Sound Detection with Self-Supervised Audio Embeddings
Mar 4, 2026Kevin Wilkinghoff, Sarthak Yadav, Zheng-Hua TanAnomaly Detection
Department of Electronic Systems, Aalborg University, Aalborg, Denmark · Pioneer Centre for Artificial Intelligence, Denmark
Relative Mismatch: Local-Reference Calibration of Feature-Space Flows for Anomalous Sound Detection
Sep 24, 2026Anbai Jiang, Xinhu Zheng, Lvxin Xu +6Anomaly Detection
Tsinghua University · Shanghai Jiao Tong University · North China Electric Power University +1

Jun 17, 2026J/K move · Enter open · S save

Scoring Backends Matter More Than Pooling: A Systematic Study of Training-Free Anomalous Sound Detection under Domain Shift

Jingwen Zhou, Mingzhe Wang

Xidian University, Xi’an, China

Distribution Shift Unsupervised Anomaly Detection Anomaly Detection

Open paper AskPDF arXiv

Training-free anomalous sound detection (ASD) scores a test clip against a memory bank of normal embeddings from a frozen pretrained audio encoder. Recent work attributes domain-shift robustness mainly to how frame-level features are pooled over time; the scoring backend applied on top of the pooled embedding has received far less systematic attention. Using a single frozen BEATs encoder on the DCASE 2023 Task 2 development set (all seven machine types), we cross four classical backends -- nearest-neighbor cosine distance, Mahalanobis distance, locally density-normalized kNN, and PCA-subspace reconstruction residual -- with three temporal poolings (mean, GeM, max). Switching the backend moves target-domain AUC by 13.8 points on average (up to 53.8), whereas switching the pooling moves it by only 3.2 points: in this training-free regime, the backend, not the pooling, dominates domain-shift robustness. No backend wins everywhere, but the machine-dependent pattern reproduces on the DCASE 2025 development data (fan, bearing). Exploiting this, we propose a label-free score fusion that z-normalizes each backend with its training-bank self-scores and takes the minimum; it reaches a harmonic-mean target AUC of 63.3% versus 64.4% for the per-machine oracle, surpassing every fixed single backend while preserving source-domain accuracy. We also report a negative result: selecting a backend by source-domain pseudo-validation with proxy outliers fails, because all backends saturate on the proxy task.

Abstract

Explore similar work

Scoring Backends Matter More Than Pooling: A Systematic Study of Training-Free Anomalous Sound Detection under Domain Shift

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

Relative Mismatch: Local-Reference Calibration of Feature-Space Flows for Anomalous Sound Detection