Training-Free Anomalous Sound Detection

• Training-free anomalous sound detection is a paradigm that uses frozen pretrained encoders to extract fixed normal embeddings, bypassing task-specific training.
• It employs classical nonparametric backends like kNN, Mahalanobis, and advanced temporal pooling methods to score anomalies using statistical measures.
• Key components include memory-bank construction, pooling strategies such as RDP, and fusion techniques that improve robustness under domain shift and first-shot monitoring.

Training-free anomalous sound detection (ASD) denotes a class of ASD systems in which a test clip is scored against a memory bank of normal embeddings produced by a frozen pretrained audio encoder. In this regime, the encoder is not fine-tuned, no classifier or anomaly detector is trained, and anomaly scoring is performed with classical nonparametric or statistical backends on fixed embeddings. The paradigm is motivated by first-shot and domain-shifted machine-condition monitoring, where only normal reference sounds are available, target-domain normal data are scarce, and task-specific training is undesirable or infeasible (Zhou et al., 17 Jun 2026, Saengthong et al., 2024).

1. Problem setting and benchmark regime

Training-free ASD sits within the broader unsupervised ASD setting, but it imposes a stricter operational constraint. In the DCASE Task 2 lineage, unsupervised ASD originally meant that no anomalous training data were provided and training used only normal sounds. DCASE 2021 formalized the domain-shifted setting, where normal training and test acoustics differ because of factors such as operating speed, machine load, viscosity, heating temperature, environmental noise, and signal-to-noise ratio, with source and target domains denoted by $\mathcal{D}_S$ and $\mathcal{D}_T$ (Kawaguchi et al., 2021). Later first-shot settings intensified this imbalance: a representative configuration uses $990$ source-domain normal clips and $10$ target-domain normal clips per machine, for a total of $N=1000$ normal training clips (Zhou et al., 17 Jun 2026).

Evaluation follows the DCASE conventions. For DCASE 2020, the official score is the arithmetic mean of AUC and pAUC with $p=0.1$. For DCASE 2021, the score is $$\mathrm{Hmean}(\mathrm{s\_auc}, \mathrm{s\_pauc}, \mathrm{t\_auc}, \mathrm{t\_pauc})$$. For DCASE 2022–2024, the score is $$\mathrm{Hmean}(\mathrm{smix\_auc}, \mathrm{tmix\_auc}, \mathrm{mix\_pauc})$$ (Fujimura et al., 14 Jul 2025). These metrics encode the practical ASD priority that low-false-positive behavior matters, particularly under source-target mismatch.

ASDKit systematizes this evaluation landscape through a unified four-step pipeline: train the frontend, extract and store frontend outputs, train the backend and compute anomaly scores, and compute official evaluation scores. Its raw-feature recipes raw_spec, raw_beats, and raw_eat are the closest toolkit instantiations of training-free ASD, because they skip task-specific frontend training and rely on frozen features plus a backend such as kNN (Fujimura et al., 14 Jul 2025). This makes the distinction between “anomaly-label-free” and genuinely “training-free” experimentally visible.

2. Canonical architecture of training-free ASD

The canonical training-free pipeline consists of three stages: fixed representation extraction, temporal pooling, and memory-bank scoring. With a frozen encoder, a clip $x$ is mapped to frame features

$F(x)\in\mathbb{R}^{T\times d},$

then a temporal pooling operator $\mathcal{D}_T$0 produces a clip embedding

$\mathcal{D}_T$1

and the memory bank is

$\mathcal{D}_T$2

At test time, a backend computes a scalar anomaly score from the test embedding against $\mathcal{D}_T$3 (Zhou et al., 17 Jun 2026).

GenRep is a prototypical realization of this design. It uses a pre-trained BEATs model as a fixed feature extractor without fine-tuning, applies temporal mean pooling only, builds source and target memory banks of normal embeddings, augments the target bank with MemMixup, and applies Domain Normalization to align source and target score scales. On the DCASE2023T2 Eval set, GenRep reports AUC Source $\mathcal{D}_T$4, AUC Target $\mathcal{D}_T$5, pAUC $\mathcal{D}_T$6, and Official Score $\mathcal{D}_T$7, with a target-domain AUC gain of $\mathcal{D}_T$8 points over the OE baseline (Saengthong et al., 2024).

Representative method	Frozen representation	Backend or central mechanism
GenRep (Saengthong et al., 2024)	BEATs	kNN, MemMixup, Domain Normalization
“Temporal Pooling Strategies for Training-Free Anomalous Sound Detection with Self-Supervised Audio Embeddings” (Wilkinghoff et al., 4 Mar 2026)	OpenL3, BEATs, EAT, Dasheng	mean, GeM, RDP, RDP + GeM
“Scoring Backends Matter More Than Pooling” (Zhou et al., 17 Jun 2026)	BEATs_iter3+	kNN cosine, Mahalanobis, locally density-normalized kNN, PCA residual, z-min fusion
BEAM (Saengthong et al., 14 Mar 2026)	handcrafted or frozen deep features	per-sub-band retrieval, uniform aggregation, AdaBEAM
“Timbre Difference Capturing in Anomalous Sound Detection” (Nishida et al., 2024)	audio embeddings	kNN detection with timbre-difference estimation

This pipeline is simple in form but not trivial in design. Once the encoder is frozen, the principal degrees of freedom move to pooling, scoring geometry, memory-bank construction, and score calibration. The later literature concentrates precisely on those components.

3. Temporal pooling as an information bottleneck

Training-free ASD almost always begins with a sequence of frame-level embeddings and therefore requires a reduction to a fixed-dimensional vector. Mean pooling has been the de facto baseline: $\mathcal{D}_T$9 Alternative fixed poolings include coordinate-wise max pooling and generalized mean pooling,

$990$0

which interpolates between mean-like and max-like behavior (Wilkinghoff et al., 4 Mar 2026).

A more recent proposal is Relative Deviation Pooling (RDP), which weights frames according to their deviation from the sequence mean. If

$990$1

then the RDP weights are

$990$2

and the pooled representation is

$990$3

The hybrid RDP + GeM strategy applies GeM under RDP-derived weights (Wilkinghoff et al., 4 Mar 2026).

Across five DCASE benchmarks and four frozen embedding models, mean pooling yields $990$4, max pooling $990$5, GWRP $990$6, GeM $990$7, RDP $990$8, and RDP + GeM $990$9; with embedding-specific tuning, RDP + GeM reaches $10$0. On individual embeddings, BEATs improves from $10$1 with mean pooling to $10$2 with RDP, and Dasheng improves from $10$3 to $10$4 with RDP (Wilkinghoff et al., 4 Mar 2026).

These results identify pooling as a real performance bottleneck. At the same time, a later controlled study using a single frozen BEATs encoder reported that switching temporal pooling moves target-domain AUC by only $10$5 points on average, whereas switching the scoring backend moves it by $10$6 points (Zhou et al., 17 Jun 2026). Taken together, the evidence indicates that pooling matters, but its importance is conditional on the embedding family and on the backend layered above it.

4. Scoring backends, geometric assumptions, and fusion

The backend specifies the geometry of normality in pooled-embedding space. In one systematic comparison, four classical backends were crossed with three poolings on the DCASE 2023 Task 2 development set: nearest-neighbor cosine distance, Mahalanobis distance, locally density-normalized $10$7NN, and PCA-subspace reconstruction residual (Zhou et al., 17 Jun 2026).

The cosine distance is

$10$8

and the nearest-neighbor cosine score is

$10$9

Mahalanobis scoring uses the bank mean $N=1000$0 and a Ledoit–Wolf shrinkage covariance $N=1000$1: $N=1000$2 Locally density-normalized $N=1000$3NN divides the $N=1000$4NN score by a local density term $N=1000$5, and PCA residual scoring treats anomalies as points outside a retained $N=1000$6-variance normal subspace (Zhou et al., 17 Jun 2026).

The empirical conclusion is explicit: backend choice dominates pooling choice in this regime. On DCASE 2023 target-domain AUC, the average range induced by changing the backend is $N=1000$7 points, compared with $N=1000$8 for changing the pooling; on DCASE 2025 target-domain AUC, the corresponding values are $N=1000$9 and $p=0.1$0. The largest target-domain backend swing is on fan, from $p=0.1$1 for PCA residual to $p=0.1$2 for density-normalized $p=0.1$3NN. No backend wins everywhere: density-normalized $p=0.1$4NN is best on target for $p=0.1$5 machines, while PCA residual beats plain $p=0.1$6NN on $p=0.1$7 machines (Zhou et al., 17 Jun 2026).

Because no single backend is universally optimal, the same study proposes a label-free fusion. For each backend $p=0.1$8, the test score is standardized using training-bank self-scores,

$p=0.1$9

and the fused score is the minimum across backends,

$$\mathrm{Hmean}(\mathrm{s\_auc}, \mathrm{s\_pauc}, \mathrm{t\_auc}, \mathrm{t\_pauc})$$0

On DCASE 2023, the harmonic mean of target-domain AUC across machines is $$\mathrm{Hmean}(\mathrm{s\_auc}, \mathrm{s\_pauc}, \mathrm{t\_auc}, \mathrm{t\_pauc})$$1 for the per-machine oracle, $$\mathrm{Hmean}(\mathrm{s\_auc}, \mathrm{s\_pauc}, \mathrm{t\_auc}, \mathrm{t\_pauc})$$2 for z-min fusion, $$\mathrm{Hmean}(\mathrm{s\_auc}, \mathrm{s\_pauc}, \mathrm{t\_auc}, \mathrm{t\_pauc})$$3 for the best fixed backend, $$\mathrm{Hmean}(\mathrm{s\_auc}, \mathrm{s\_pauc}, \mathrm{t\_auc}, \mathrm{t\_pauc})$$4 for the expected blind pick, and $$\mathrm{Hmean}(\mathrm{s\_auc}, \mathrm{s\_pauc}, \mathrm{t\_auc}, \mathrm{t\_pauc})$$5 for pseudo-validation selection. A negative result is equally significant: pseudo-validation with proxy outliers fails because pseudo-AUC saturates at about $$\mathrm{Hmean}(\mathrm{s\_auc}, \mathrm{s\_pauc}, \mathrm{t\_auc}, \mathrm{t\_pauc})$$6 to $$\mathrm{Hmean}(\mathrm{s\_auc}, \mathrm{s\_pauc}, \mathrm{t\_auc}, \mathrm{t\_pauc})$$7 for all backends, so the selection criterion cannot distinguish them (Zhou et al., 17 Jun 2026).

5. Localized retrieval, sub-band scoring, and interpretable extensions

Global whole-clip nearest-neighbor matching assumes that one reference embedding is appropriate for all frequency regions. BEAM challenges that assumption by partitioning a clip vector into $$\mathrm{Hmean}(\mathrm{s\_auc}, \mathrm{s\_pauc}, \mathrm{t\_auc}, \mathrm{t\_pauc})$$8 contiguous sub-bands and storing one memory bank per band: $$\mathrm{Hmean}(\mathrm{s\_auc}, \mathrm{s\_pauc}, \mathrm{t\_auc}, \mathrm{t\_pauc})$$9 For a query band $$\mathrm{Hmean}(\mathrm{smix\_auc}, \mathrm{tmix\_auc}, \mathrm{mix\_pauc})$$0, the local score is

$$\mathrm{Hmean}(\mathrm{smix\_auc}, \mathrm{tmix\_auc}, \mathrm{mix\_pauc})$$1

with cosine distance

$$\mathrm{Hmean}(\mathrm{smix\_auc}, \mathrm{tmix\_auc}, \mathrm{mix\_pauc})$$2

and the final score is the uniform average

$$\mathrm{Hmean}(\mathrm{smix\_auc}, \mathrm{tmix\_auc}, \mathrm{mix\_pauc})$$3

This replaces one global neighbor with one best neighbor per band and removes the energy-coupled aggregation of global cosine matching (Saengthong et al., 14 Mar 2026).

BEAM is supported by both analysis and benchmark results. The paper derives a normal-variance bound of the form

$$\mathrm{Hmean}(\mathrm{smix\_auc}, \mathrm{tmix\_auc}, \mathrm{mix\_pauc})$$4

and reports that sub-band matching reduces normal-score variance while often preserving the anomaly-normal mean gap. Empirically, on DCASE2020 sec-dev with Log-Mel, global matching yields $$\mathrm{Hmean}(\mathrm{smix\_auc}, \mathrm{tmix\_auc}, \mathrm{mix\_pauc})$$5, BEAM $$\mathrm{Hmean}(\mathrm{smix\_auc}, \mathrm{tmix\_auc}, \mathrm{mix\_pauc})$$6, and AdaBEAM $$\mathrm{Hmean}(\mathrm{smix\_auc}, \mathrm{tmix\_auc}, \mathrm{mix\_pauc})$$7; with BEATs iter3 on DCASE2020 sec-eval, the sequence is $$\mathrm{Hmean}(\mathrm{smix\_auc}, \mathrm{tmix\_auc}, \mathrm{mix\_pauc})$$8 (Saengthong et al., 14 Mar 2026).

A different extension of the same local-neighborhood principle appears in “Timbre Difference Capturing in Anomalous Sound Detection.” There, the $$\mathrm{Hmean}(\mathrm{smix\_auc}, \mathrm{tmix\_auc}, \mathrm{mix\_pauc})$$9-nearest normal neighbors in embedding space are used not only for anomaly scoring but also for explanation in predefined timbre attributes—sharpness, roughness, boominess, brightness, and depth—using objective metrics from psychoacoustical timbral models. No anomalous sounds are required for training the explanation mechanism, and BEATs gives the best overall MAE for timbre-difference capturing on the MIMII DG evaluation setup (Nishida et al., 2024). This suggests that training-free ASD can support explanation as well as detection when comparison is localized to condition-matched normal neighbors.

Robustness to mixture shift remains a separate issue. “Retaining Mixture Representations for Domain Generalized Anomalous Sound Detection” argues that frozen SSL encoders can break down on low-SNR mixtures of machine and noise, and proposes a retain-not-denoise pretraining strategy with a multi-label tagging loss and a mixture alignment loss. The method is not training-free end-to-end, but it is designed to improve a downstream training-free $x$0-nearest-neighbor regime by preserving machine-plus-noise information in the frozen representation (Saengthong et al., 29 Oct 2025).

6. Taxonomy boundaries and common misconceptions

A central misconception in the ASD literature is that “unsupervised,” “normal-only trained,” “parameter-free component,” and “training-free” are interchangeable. They are not. Training-free ASD, in the strict sense used by recent frozen-embedding work, requires a frozen pretrained encoder, no fine-tuning, no trained classifier or anomaly detector, and a classical scoring rule over a memory bank (Zhou et al., 17 Jun 2026).

Many recent ASD methods are explicitly outside that category. Noisy-ArcMix is a trained ASD model that uses ArcFace-style additive angular margin loss, mixup, Noisy-ArcMix, and TAgram; it reports Average AUC $x$1, Average pAUC $x$2, and mAUC $x$3 on DCASE 2020, but it is not training-free (Choi et al., 2023). ASD-Diffusion trains a U-Net denoiser on normal FBank features and uses diffusion reconstruction plus an anomaly filter; it reports overall hmean $x$4 on the DCASE 2023 development set, but it likewise requires normal-data training (Zhang et al., 2024). IDC-TransAE and GRLNet are normal-only trained detectors whose inference is training-free only in the loose sense that no test-time adaptation is performed (Guan et al., 2023, Sha et al., 2022).

The same boundary applies to “parameter-free” or “low-training” systems. The modified-FBank plus SimAM approach uses a parameter-free attention module, but the EAT backbone is still fine-tuned and an ArcFace classifier is trained (Zhong et al., 21 Aug 2025). Large pretrained models adapted with LoRA, machine-aware group adapters, and dual-level contrastive learning are parameter-efficient transfer systems rather than training-free systems (Han et al., 17 Aug 2025). First-shot synthetic-anomaly methods based on AudioLDM and TWFR-GMM remain training-light rather than training-free, because the generator is fine-tuned and the detector is calibrated on synthetic data (Zhang et al., 2023). Serial-OE is further away still, because it uses outlier exposure, a trained feature extractor, and GMMs, and can incorporate small amounts of real anomalous data (Kuroyanagi et al., 25 May 2025).

The practical implication is that training-free ASD is best understood as a backend-centric paradigm rather than as a synonym for anomaly-label-free learning. Its distinguishing research questions concern frozen representations, pooling, scoring geometry, score normalization, memory-bank design, and robustness under domain shift. The recent literature indicates that substantial gains can be obtained in that regime without task-specific training, but it also makes clear that adjacent trained methods remain strong baselines and should not be conflated with the stricter training-free category (Saengthong et al., 2024, Wilkinghoff et al., 4 Mar 2026, Zhou et al., 17 Jun 2026).