Replay-attack detectors try to tell live speech from speech played back through a loudspeaker. Their best cue is often the group delay, meaning the frequency derivative of the phase spectrum, which shows how long each frequency takes to pass through the recording chain. Group delay is fragile in a specific way: the function blows up wherever the spectrum has a near-zero, and additive noise moves those near-zeros around. In the ASVspoof 2019 challenge, systems that did well on simulated replay dropped sharply on real room recordings. We propose to stop reading group delay only on the unit circle, which is where those zeros sit. Instead we read it on a family of circles of radius $r$ in the complex plane and treat $r$ as a scale that "runs," much as a coupling runs with energy in the renormalization group, the physicist's method for tracking how parameters change with scale.
Along this flow there is an exact identity: the rate at which the average log-magnitude changes with $\log r$ equals minus the average group delay. That rate is a step function that jumps only when $r$ crosses the radius of a spectral zero. It therefore behaves exactly like a beta function that changes its coefficient each time a heavy particle is integrated out. We hypothesize that the replay channel's signature lives in this radial staircase of zero radii, and that additive noise mainly scrambles the steps nearest $r=1$. A detector that sees the whole flow should then lose less on real recordings than one that sees only the unit-circle group delay.
Core Claim
Take one analysis frame $x[0],\dots,x[N-1]$ with z-transform $X(z)=\sum_n x[n]z^{-n}$. Define $M(r)=\frac{1}{2\pi}\int_0^{2\pi}\log|X(re^{i\theta})|\,d\theta$ as the mean log-magnitude on the circle of radius $r$, and $\langle\tau_g\rangle_r$ as the mean group delay on that circle. Jensen's formula from complex analysis, applied to the polynomial $z^{N-1}X(z)$, together with the argument principle gives
$$\frac{dM}{d\log r} \;=\; N_{\rm in}(r)-(N-1) \;=\; -\langle \tau_g\rangle_r ,$$
where $N_{\rm in}(r)$ is the number of zeros of $X$ with modulus below $r$. The identity holds whenever no zero lies exactly on the circle.
$N_{\rm in}(r)$ is piecewise constant. Its jumps sit at the zero radii, which are the "thresholds" of this flow. Between thresholds, $M$ is exactly linear in $\log r$. This is the same structure as the step-function match-and-run procedure used in top-down particle models such as G_2-MSSM: couplings run between mass thresholds with a fixed beta coefficient, and the coefficient changes as particles decouple. The analogy supplies a concrete design choice: describe the signal by the full staircase rather than by its value at a single scale.
A second lens comes from moduli-dependent particle spectra in string compactifications. There, the net number of generations is a topological index that stays fixed under deformation. The individual cohomology dimensions, by contrast, jump at special loci of moduli space (Vector Bundle Extensions, Sheaf Cohomology, and the Heterotic Standard Model; Heterotic compactification, an algorithmic approach). The frame polynomial has the same split. Its total zero count is fixed at $N-1$, a coarse invariant. How those zeros are distributed across radii, and especially which side of $r=1$ they fall on, is what jumps as the acoustic channel or the noise changes.
Our claim follows from this picture. Replay through a loudspeaker and a room adds non-minimum-phase zeros, meaning zeros outside the unit circle. The fine positions of zeros near $r=1$ are noise-dominated, while the coarse counts are more stable. A detector should therefore use counts away from $r=1$ to stay robust and use the near-circle structure only where it is reliable. These physics papers act as a lens. We reuse no code or equations from them beyond the analogy.
Relation to Prior Work
The group delay papers supply the feature family we generalize. The modified group delay feature (MODGDF) tames spikes at spectral zeros by replacing the magnitude with a smoothed envelope and adding exponents $\alpha,\gamma$ that need tuning for each domain. Our radius $r$ plays the same regularizing role, pulling the evaluation contour away from zeros, but it does so in a principled one-parameter family whose behavior is described exactly by the identity above.
The DKU replay system showed that the group delay gram (a time-frequency image of group delay) fed to a ResNet was the strongest single feature for physical-access replay detection. That system evaluates only at $r=1$. Detecting Converted Speech and the ASVspoof overview papers establish that phase carries cues that magnitude features miss.
ASVspoof 2019: spoofing countermeasures motivates the proposal. It documents the gap between simulated and real replay, and it reports that fusing magnitude and phase systems gives diminishing returns on physical access. That second finding constrains us: simply adding another channel to a fusion is unlikely to help.
From the renormalization group side, Wavelet Conditional Renormalization Group and RINO show that organizing a representation by scale helps optimization and transfer from simulation to real data. RINO is the closest methodological cousin, since it enforces invariance across jet-clustering scales to reduce the domain gap between simulated and real collider data. Our flow parameter is a contour radius rather than a spatial or momentum scale. It also comes with an exact conservation-like identity, which RINO's clustering scales do not have.
Research Program
The front end evaluates each frame on a grid of roughly 16 to 32 radii in $[0.85, 1.15]$. This is done by exponentially weighting the frame, $x[n]r^{-n}$, followed by a fast Fourier transform. Each frame yields two outputs. The first is a radius-frequency group delay image, the "running group delay gram." The second is the staircase $N_{\rm in}(r)$. We estimate the staircase without root-finding, by finite differences of $M(r)$, and check it against explicit polynomial roots on a subset of frames.
The staircase stacked across frames becomes a time-radius image. Its cell just outside $r=1$ directly measures how many maximum-phase zeros each frame has. Both images feed the same ResNet-34 with global average pooling that the DKU system uses. Keeping the classifier fixed means any gain is attributable to the front end.
We then add one learned ingredient borrowed from the renormalization group. Inspired by RINO, a consistency loss asks embeddings computed from the outer radii to agree across noise-augmented copies of an utterance. Embeddings from near $r=1$ are left free. This tests whether explicitly separating the "infrared" (coarse, far from the circle) part of the flow from the "ultraviolet" (fine, near the circle) part matters beyond simply having multi-radius input.
Before any benchmark training, we run a mechanistic study on controlled simulations. We convolve clean speech with measured loudspeaker and room impulse responses at known configurations. For each configuration we measure two things. First, how the zero-radius distribution shifts between live and replayed speech. Second, how stable each radius slice is under added white and babble noise across a range of signal-to-noise ratios (SNRs), quantified as the correlation between clean and noisy features. The hypothesis fails at this stage if noise destroys the outer-radius slices as fast as the $r=1$ slice.
Baselines
The naive reference point is the linear frequency cepstral coefficient (LFCC) front end with the same ResNet. On physical-access development data, the DKU system reports a development equal error rate (EER) of 11.96% for the classical LFCC pipeline, falling to 2.62% with the ResNet. We rerun this baseline with our data splits and training budget.
The strongest single-feature competitor is the DKU group delay gram with data augmentation. It reports a development EER of 1.03%, and its six-model fusion reaches an evaluation EER of 0.66%. It is the strongest competitor because it is the same phase-derivative idea evaluated at $r=1$ only, so it isolates exactly what our flow adds. We reimplement it with an identical network, augmentation, number of epochs and hyperparameter search budget. We report single-system results against our single system and fused results against a fusion with the same number of members.
The challenge-level reference numbers from the ASVspoof 2019 results paper are a best physical-access minimum tandem detection cost (min t-DCF) of 0.1437 at 0.39% EER, and a best min t-DCF of 0.3855 on the real replay recordings.
A MODGDF baseline, with $\alpha,\gamma$ tuned on development data under the same tuning budget, tests whether hand-tuned smoothing already captures what the flow offers.
The key ablation is a multi-resolution stack of MODGDF and group delay grams at several smoothing levels. It has the same number of input channels as our radial stack, which rules out "more channels" as the explanation. Two further ablations remove the staircase image and remove the consistency loss, respectively.
Beating the baselines means two things at matched single-system budget. First, matching the DKU group delay gram on simulated physical-access evaluation, within statistical error. Second, a clearly lower min t-DCF on the real replay subset, and a smaller ratio of real to simulated t-DCF, than every baseline including the matched-channel ablation.
Why not just…?
Why not just tune MODGDF's $\alpha$ and $\gamma$, or stack several settings? This is the most serious threat. MODGDF already suppresses zero-induced spikes, and the MODGDF paper reports 98.5% speaker recognition on clean TIMIT but 66% on noisy telephone speech (NTIMIT), which shows its smoothing does not remove noise sensitivity. Smoothing the magnitude envelope, however, does not separate zeros by which side of the unit circle they lie on, and that inside/outside split is our hypothesized replay cue. The matched-channel multi-resolution MODGDF ablation settles the question directly. If it closes the gap to our system on real replay, the radial flow adds nothing, and we will report that.
Why not just add more noise augmentation? The DKU system already relies heavily on augmentation. Its gains on simulated data did not stop the challenge-wide drop on real recordings documented in the ASVspoof 2019 results. We give every baseline the same augmentation pipeline. If augmentation alone explains robustness, the baselines will improve as much as our system does, and the gap will vanish.
Why not just fuse stronger magnitude features? On physical access, the challenge paper reports that score-level fusion of magnitude and phase systems yields diminishing returns. That suggests the missing information is not recovered by more magnitude variety. We nonetheless include a fusion of LFCC and the group delay gram with as many members as our fusion.
Why not learn the front end from raw waveforms? In principle a learned filterbank could discover radial weighting. The research file contains no reported result on Lacuna for raw-waveform front ends on the real replay subset, so we cannot cite a number. We treat this as an extension: we would initialize a learnable exponential weighting at our radius grid and test whether learning moves it.
Proposed Experiments
Training uses the ASVspoof 2019 database physical-access training data. Evaluation covers the simulated physical-access evaluation set and its real replay recordings. Metrics are min t-DCF and EER, with confidence intervals bootstrapped over speakers.
We add a controlled noise sweep in which babble and stationary noise are added to evaluation audio from 30 dB down to 0 dB SNR. The research file contains no reported result on Lacuna for any baseline under such a sweep, so all numbers there will come from our own reimplementations. As a secondary check that the flow does not hurt on synthetic speech, we evaluate on logical access, where the best reported system reaches min t-DCF 0.0692 and 0.22% EER (ASVspoof 2019 results).
The decisive checks are three. The first is the mechanistic simulation study: outer-radius features must be measurably more stable under noise than $r=1$ group delay while still separating live from replayed speech. The second is the real-replay comparison against the matched-channel MODGDF stack. The third is a per-radius attribution study, in which we zero out one radial slice at a time. If replay evidence concentrates in the slice just outside $r=1$ and survives noise, the staircase picture is supported. If the network relies only on the slice at $r=1$, the flow is a reparameterization and nothing more.
References
- <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:msub><mml:mi>G</mml:mi><mml:mn>2</mml:mn></mml:msub></mml:math>-MSSM: An<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi>M</mml:mi></mml:math>theory motivated model of particle physics
- Vector bundle extensions, sheaf cohomology, and the heterotic standard model
- Heterotic compactification, an algorithmic approach
- The modified group delay feature: a new spectral representation of speech
- The DKU Replay Detection System for the ASVspoof 2019 Challenge: On Data Augmentation, Feature Representation, Classification, and Fusion
- Detecting Converted Speech and Natural Speech for anti-Spoofing Attack in Speaker Recognition
- ASVspoof 2019: Spoofing Countermeasures for the Detection of Synthesized, Converted and Replayed Speech
- Wavelet Conditional Renormalization Group
- RINO: Renormalization Group Invariance with No Labels
- The ASVspoof 2019 database