A microphone watching one machine on a factory floor also hears the machines next to it. Those neighbors do not switch on and off at random from one fraction of a second to the next. They run in duty cycles set by controllers and production queues, so when a neighbor is loud now, it is very likely still loud a moment later. Today's best training-free anomaly detectors for machine sound do not use this fact. They summarize a clip by up-weighting whichever frames deviate most from the clip's average, and they treat every short deviation as a possible fault. A burst from a neighboring valve looks to them exactly like a cracked bearing.
This note proposes to borrow a lesson from wireless networks. When interference from nearby transmitters persists across time slots, models that treat each slot as independent badly misjudge reliability. Failures arrive in streaks, and the retransmission gains that the independent model predicts never appear. In machine-sound monitoring, the analogous step is to treat the unlabeled deployment stream as the evidence for a persistent hidden interferer state. We would infer that state from the audio alone, without controller logs, and score anomalies only after discounting the frames the interferer explains. The two-sentence version is this. A neighbor's interference shows up as a recurring shift in the embedding that sticks for many frames and comes back repeatedly, while a fault is a rare deviation that this sticky recurring shift does not explain. Modeling that stickiness explicitly should cut false alarms whenever neighbors are bursty, and the gain should grow with how long their bursts last.
Core Claim
The claim is specific and falsifiable. Take a frozen audio foundation model, meaning a large network pre-trained on general audio whose frame-level embeddings we use without retraining. Add a small two-state switching model fit online to the unlabeled stream from the deployed microphone: a hidden "interferer off / interferer on" variable that tends to stay in its current state. Correcting embeddings by the inferred state before pooling and scoring will beat the strongest training-free detector under neighbor interference. The improvement will come specifically from the temporal persistence term. An otherwise identical model without persistence, a plain two-component mixture that treats frames independently, should recover little of the gain. The gap between the two should widen as the interferer's mean burst length grows, and it should vanish when the interferer toggles independently from frame to frame.
The motivating evidence comes from wireless analysis. A Unified Stochastic Geometry Model for MIMO Cellular Networks With Retransmissions models multi-antenna cellular links with retransmissions and shows that ignoring the correlation between attempts gives up to a 3 dB error in predicted signal-to-interference gains. Capacity of Cellular Wireless Network shows that receivers stuck in high-interference states see heavy-tailed delays instead of exponential ones. Stochastic Geometry Analysis of Spatial-Temporal Performance in Wireless Networks: A Tutorial attributes this persistence to two causes, fixed geometry and queue-driven transmitter activity, and introduces the meta distribution, which looks at the distribution of per-receiver success rates rather than a network-wide average. We take the structural lesson from these papers and do not use their closed-form point-process mathematics. Our acoustic interferers are a handful of known neighbors, not a random spatial field, so stochastic geometry would add formalism without predictive value.
What Prior Acoustic Work Does and Leaves Open
The direct target is Temporal Pooling Strategies for Training-Free Anomalous Sound Detection with Self-Supervised Audio Embeddings. It introduces Relative Deviation Pooling (RDP), which weights each frame by its distance from the clip mean. It also combines RDP with a generalized mean, a pooling rule with a tunable exponent that interpolates between averaging and taking the maximum. With a BEATs backbone, this reaches a 63.2% harmonic-mean AUC on the DCASE 2025 anomaly detection challenge, where AUC is the area under the curve of detection rate against false-alarm rate. RDP amplifies any deviation, which is exactly why it is vulnerable to bursty neighbors. We keep its pooling and change what counts as a deviation.
The closest attempt to handle interference is Activity-Guided Industrial Anomalous Sound Detection against Interferences. It feeds binary on/off logs from programmable logic controllers into a source-separation network and then runs an autoencoder detector. On MIMII valve mixtures it raises AUC from 0.626 to 0.751. It differs from our proposal in three ways. It needs the logs, it needs isolated recordings of each machine to train the separator, and its authors state it cannot handle always-on machines. Our method infers activity from the stream itself, needs no isolated audio, and treats an always-on neighbor whose loudness fluctuates as a slowly varying state rather than a binary one. We therefore treat this paper as an oracle-style upper reference rather than a competitor on equal footing.
Two speech test-time adaptation papers supply pieces of the method and also show what is missing. E-BATS finds that acoustic shifts in foundation-model embeddings act mostly as translations of the mean rather than changes in covariance. That finding is what licenses our model of the interferer as an additive offset in embedding space. E-BATS also stabilizes streaming adaptation with an exponential moving average whose smoothing constant is fixed by hand. Our persistence parameter is the principled counterpart: it is estimated from the stream and corresponds to a mean burst length. Advancing Test-Time Adaptation for Acoustic Foundation Models in Open-World Shifts regularizes predictions to be smooth across neighboring frames. That is smoothness of the target signal, not a model of the external interferer, and it requires backpropagation at test time, which we avoid. Finally, Handling Domain Shifts for Anomalous Sound Detection: A Review of DCASE-Related Work names the underlying conundrum: a detector trained only on normal sound cannot tell a benign new condition from a fault. Our proposal addresses one physically common instance of that conundrum. When a shift comes from a neighbor, temporal persistence and recurrence are the cues that separate it from a fault.
Research Program
The detector has three parts. Frame embeddings $z_t$ come from frozen BEATs, chosen because AnoPatch shows general-audio models beat speech models on machine sound. On the unlabeled deployment stream, which spans many consecutive clips rather than one, we fit a sticky two-state switching model:
$$z_t = \mu + s_t\, d + \varepsilon_t,\qquad \Pr(s_t = s_{t-1}) = \rho,$$
Here $s_t \in \{0,1\}$ is the hidden interferer state, $\mu$ is the target-only mean, $d$ is the interferer's offset direction in embedding space, $\varepsilon_t$ is residual variation, and $\rho$ is the probability of staying in the current state. A stay probability of $\rho$ means a mean burst length of $1/(1-\rho)$ frames, the standard geometric dwell time of a two-state Markov chain. Online forward filtering gives the posterior $\gamma_t = \Pr(s_t=1 \mid z_{1:t})$. For always-on neighbors we replace the binary state with a continuous loudness state that follows a first-order autoregressive process, which keeps the same persistence logic.
The anomaly score uses the interference-corrected frames $\tilde z_t = z_t - \gamma_t d$. We pool them with the unchanged RDP plus generalized-mean rule and score the pooled vector by distance to the nearest neighbor in a memory bank of normal training clips. The frames are corrected but not discarded, because a fault that happens during a neighbor burst must still be detectable.
To stop the switching model from absorbing real faults, we impose two constraints. The offset $d$ must recur, meaning it must explain deviations across many clips in the stream. Its fitted occupancy must also exceed a minimum fraction of frames. A one-off fault cannot meet either requirement. Testing whether these constraints hold is part of the study. The method adds a two-state filter per frame and requires no gradient steps, so it can run on the edge devices where E-BATS was motivated.
The analytical contribution is a set of evaluation lenses borrowed from wireless systems. Alongside AUC and partial AUC, we report the distribution of false-alarm run lengths and a per-stream false-alarm meta distribution: the spread of false-alarm rates across individual deployments rather than their mean. The wireless theory predicts that persistent interference produces a heavy tail of "unlucky" deployments. A method that only improves the average can leave that tail untouched, and it is exactly the tail that makes operators stop trusting alarms.
Baselines
The naive reference is BEATs embeddings with mean pooling and nearest-neighbor scoring. RDP shows this washes out transient faults, but it is also the method least distracted by bursts, so it bounds the problem from the other side.
The strongest existing method is BEATs with RDP plus generalized-mean pooling from Temporal Pooling Strategies, which reports 63.2% harmonic-mean AUC on DCASE 2025. We will run it with the same backbone, the same frame rate, the same memory bank and the same distance. Its pooling exponent and our persistence prior will each get an identical hyperparameter search budget, tuned on development streams only. Beating it means two things. On our interference streams we must win on AUC and partial AUC, with confidence intervals from resampling streams. On clean DCASE 2025 we must stay within noise of 63.2%, because a method that helps under interference but hurts clean performance has not won.
The second learned reference is AnoPatch, which fine-tunes BEATs with metadata supervision and reports 68.87 mean AUC on DCASE 2023. We run it from released settings to show whether supervised fine-tuning already absorbs neighbor interference.
The upper reference is the controller-log-guided pipeline of Activity-Guided Industrial Anomalous Sound Detection, which reports 0.751 AUC versus 0.626 on MIMII valve mixtures. We reproduce its mixture protocol so the numbers are comparable. In addition, we give our own detector the true activity sequence in place of $\gamma_t$, which measures how much of the oracle gap state inference closes.
The key ablation sets $\rho$ to the state's marginal frequency. This turns the model into an independent-frame mixture with the same offset $d$, the same correction and the same pooling. Any advantage of the full model over this ablation is attributable to temporal persistence alone.
Why not just…?
Why not just augment the normal memory bank with neighbor sounds and score frames independently? This is the most serious threat. If neighbor bursts at training time resemble those at test time, frame-level nearest-neighbor scoring against an augmented bank may simply call burst frames normal. We will run this alternative with the same neighbor recordings that we use to build the test streams, mixed at random independent timings. The proposal only survives if it wins when the test-time neighbor differs from the augmentation neighbor in machine unit, gain or duty cycle, which is the domain-shift case. If augmentation matches us even there, we will report that persistence modeling is unnecessary and recommend augmentation.
Why not just smooth the anomaly scores over time with a moving average? Smoothing makes persistent interference worse. A multi-frame burst survives a moving average while a short fault is diluted, which is the opposite of what we want. A minimum filter has the reverse problem and suppresses short faults. We include both filters, with window length tuned under the same budget. The run-length analysis will show whether either one shortens false-alarm streaks without losing detection of short faults.
Why not just use the controller-log separator, or a fixed exponential moving average as in E-BATS? The separator needs logs and isolated recordings, and its authors exclude always-on machines. It serves as our oracle reference, not as a substitute. A fixed moving average is the right comparison for the persistence term. We will sweep its constant and test whether any single value matches the persistence parameter estimated per stream across interferers with different burst lengths. If one fixed constant does as well, the estimation step adds nothing.
Proposed Experiments
Public datasets lack long streams with controlled neighbor activity, so we construct semi-synthetic streams. Following the mixture protocol of the activity-guided paper, we take MIMII valve, slider, pump and fan clips. For each target machine we build hour-long streams from its normal and anomalous clips. We overlay neighbor-machine audio whose activity follows a two-state Markov chain spanning clip boundaries, with mean burst lengths swept from one frame (independent) to several clips. For fans and pumps the neighbor is always on with an autoregressive loudness envelope. We vary the target-to-interferer ratio. The test-time neighbor is a different machine unit from any neighbor seen in development. We also report the unmodified DCASE 2025 evaluation as a clean-condition check, using DCASE 2025 and MIMII as they appear in the cited papers; no reported result found on Lacuna for any method on long-stream interference benchmarks of this kind.
The decisive check is a curve: the AUC gain of the full model over the independent-frame ablation, plotted against the interferer's mean burst length. The hypothesis predicts a gain near zero at burst length one that rises with persistence. A flat curve falsifies the core claim even if the full model beats RDP, because that would mean the gain comes from the offset correction and not from persistence. The second check is the meta distribution of false alarms across streams. We expect RDP to show a heavy tail of persistently false-alarming streams, and we expect the full model to shrink that tail, not just the mean. The third check targets the risk that the switching model absorbs faults. We inject long-lasting anomalies, such as sustained faulty-valve clips spanning many frames, and measure whether the posterior $\gamma_t$ wrongly assigns them to the interferer state. If it does, the recurrence and occupancy constraints are insufficient, and we will report this as the method's boundary condition.
References
- A Unified Stochastic Geometry Model for MIMO Cellular Networks With Retransmissions
- Capacity of cellular wireless network
- Stochastic Geometry Analysis of Spatial-Temporal Performance in Wireless Networks: A Tutorial
- Temporal Pooling Strategies for Training-Free Anomalous Sound Detection with Self-Supervised Audio Embeddings
- Activity-Guided Industrial Anomalous Sound Detection against Interferences
- E-BATS: Efficient Backpropagation-Free Test-Time Adaptation for Speech Foundation Models
- Advancing Test-Time Adaptation for Acoustic Foundation Models in Open-World Shifts
- Handling Domain Shifts for Anomalous Sound Detection: A Review of DCASE-Related Work
- AnoPatch: Towards Better Consistency in Machine Anomalous Sound Detection