Speaker verification systems decide whether two recordings come from the same person. They do this by comparing a similarity score against one fixed threshold. When a test recording arrives over a narrowband phone line, a radio link, or a bone-conduction earbud, the score distribution moves. A threshold tuned on wideband audio then accepts impostors at a different rate on each channel. Precision physics has a name for this kind of error. It is a residual systematic uncertainty: a bias from imperfect knowledge of the measuring apparatus that does not shrink as more data is collected. Physics also has a tool for it. "Learning to pivot" trains a network alongside an adversary that tries to guess the nuisance condition from the network's output. Training stops the network from leaking that condition, so the output distribution no longer depends on it.
This note proposes treating the transmission channel as a continuous nuisance parameter and pivoting the verification score, not the audio or the embedding. Concretely, the distribution of impostor-trial and same-speaker-trial scores is made independent of channel parameters such as cutoff frequency, filter roll-off and noise level. The question that decides whether this is worth doing is narrow. Bandwidth can be estimated directly from the spectrum, so a simple "estimate the cutoff, then recalibrate" pipeline is a serious competitor. The proposal is built around testing whether invariance learned in simulation holds up on real radio hardware, where the channel is more than a cutoff.
Core Claim
The operational failure under limited bandwidth is drift in the score scale, as much as lost information. The most accurate per-channel embedding is therefore not the target. A model whose scores at a single global threshold give the same false-accept rate across channels is worth more. That model should also beat both plain augmentation and post-hoc cutoff-conditioned calibration. It should do so on channels whose distortions cannot be summarized by one measurable number.
Several papers motivate this claim. Robust Channel Learning for Large-Scale Radio Speaker Verification applies band-limiting augmentation and fine-tunes only the early layers. It reports that bandwidth limits cause roughly half the radio performance loss, and that channel drift concentrates in early convolutional layers. Its evaluation uses per-condition EER (equal error rate, the error level where false accepts equal false rejects). That metric hides cross-channel threshold shift. Real-time Speech Frequency Bandwidth Extension shows that models trained on one fixed cutoff break when the cutoff varies. It fixes this by randomizing the cutoff during training, which buys average robustness but no independence guarantee.
Learning to Pivot with Adversarial Networks supplies the method. It uses a mixture-density adversary, a network that outputs a full probability distribution over the nuisance value. The paper was demonstrated only on particle-physics classification. It also left the choice of the trade-off weight $\lambda$ to a downstream physics metric.
We reuse that machinery with two changes. First, the adversary acts on scores conditional on trial type, because same-speaker and impostor scores should each be channel-invariant without being merged. Second, $\lambda$ is chosen with a verification-native downstream objective that plays the role the discovery-significance metric plays in the pivot paper.
Research Program
The base system is an ECAPA-TDNN speaker encoder with cosine scoring, as used by the CRSL paper. ECAPA-TDNN is a standard convolutional speaker-embedding network. Training audio passes through a simulated channel with a parameter vector $Z$. $Z$ contains the Butterworth cutoff frequency, the filter order, and the SNR of added noise, which is the same parametric family as CRSL's BandNoiseAugment.
Training batches are built as trials. Each trial pairs a clean wideband enrollment with a channel-degraded test segment, giving a score $s$ and a label $y$ (same speaker or not). An adversary $r$ receives $s$ and $y$ and models the conditional density $p(Z \mid s, y)$. The encoder $f$ minimizes
$$ \mathcal{L}(f) = \mathcal{L}_{\text{spk}}(f) - \lambda \, \mathbb{E}\big[\log p_r(Z \mid s, y)\big], $$
where $\mathcal{L}_{\text{spk}}$ is the usual speaker-classification loss and $\lambda \ge 0$ sets the invariance strength. The adversary maximizes the log-likelihood term.
We also run a depth study that directly tests the tension between CRSL and the pivot paper. CRSL claims nuisance lives in early layers. The pivot argument says unpenalized layers pass nuisance information upward. We attach the same adversary to three places: early-layer pooled features, the final embedding, or the score. We then measure where the channel information actually survives, using the layer-wise Wasserstein drift diagnostic from CRSL.
We choose $\lambda$ by minimizing a pooled-threshold error. This is the EER computed with one threshold over trials from all channel conditions together. We report it alongside the spread of false-accept rates across conditions at that threshold. This spread is our measure of the residual systematic.
Baselines
The naive reference is ECAPA-TDNN trained on clean audio and scored with a global threshold. On narrowband FM, CRSL reports an EER of 8.99% on Vox1-N for its baseline. The strongest existing method is CRSL itself, at 4.32% EER on Vox1-N. It will be re-trained in our codebase with the same training data, the same simulated channel family, and matched GPU hours. A second augmentation baseline uses SEANet-style variable-band training, meaning cutoff randomization only. Each method gets the same hyperparameter search budget, set by a fixed number of trials.
The strongest simple competitor is cutoff-conditioned calibration. A small estimator predicts $\hat Z$ from the test spectrum, and an affine score calibration conditioned on $\hat Z$ is fitted on held-out simulated trials. It is applied on top of the CRSL model. Beating it means a lower pooled-threshold EER and a smaller false-accept-rate spread on the hardware sets (Vox1-NH, Vox1-WH). Per-condition EER must also come within about 0.3 points absolute of CRSL. Parity on simulated Vox1-N alone would not count as a win.
The ablations isolate the key ingredient. One drops the label conditioning, so the adversary sees $s$ only. The other moves the adversary to the embedding or to early layers. Together they show whether score-level, label-conditional invariance is what matters.
Why not just…?
Why not just augment with variable bandwidth, as CRSL and SEANet do? Augmentation lowers average error. It does not stop scores from shifting with the channel, and the pivot paper documents exactly this gap for unpenalized classifiers. Our pooled-threshold metric measures that shift directly. If CRSL already shows a small false-accept-rate spread, this proposal has no reason to exist. That is the first number we will check.
Why not just estimate the cutoff and recalibrate? This is the real threat. The cutoff is observable, unlike most physics nuisances. On simulated Butterworth channels we expect this baseline to match pivoting. The case for pivoting rests on hardware radio and other non-parametric distortions, where a single estimated cutoff is an incomplete description. If calibration matches pivoting on Vox1-NH and Vox1-WH too, we will report that a measured nuisance is enough and drop the adversary.
Why not just put a bandwidth-extension front-end before the verifier? French Listening Tests for the Assessment of Intelligibility, Quality, and Identity of Body-Conducted Speech Enhancement found that the EBEN extension model raised perceived quality by about 20%. It also cut identification of female speakers on throat microphones (-6% mean, -11% median). Synthesized high frequencies therefore add a new systematic of their own. We will still run SEANet-extension followed by the clean verifier as a baseline.
Proposed Experiments
Training uses VoxCeleb1 audio passed through the simulated channel family. Testing uses CRSL's simulated sets (Vox1-N, Vox1-W) and its HackRF/GNU Radio hardware sets (Vox1-NH, Vox1-WH). In the hardware sets, real transmission introduces distortions outside the simulated $Z$.
We report four quantities: per-condition EER, pooled-threshold EER, the spread of false-accept rates across conditions, and adversary log-likelihood as a leakage probe. The decisive check is the comparison of the pivoted model against cutoff-conditioned calibration on the hardware sets. A second check asks whether channel information that the early-layer adversary removes reappears at the score. That would support the pivot paper's argument over CRSL's localization claim.
As a stress test, we treat sensor type in the Vibravox corpus (forehead accelerometer, in-ear microphone, throat microphone) as a categorical nuisance. We score with ECAPA2, which the listening-test paper found tracks human identity judgments (ρ = .90). This tells us whether pivoting reduces the identity loss observed for female throat-microphone speakers. The physics benchmarks (Daya Bay, DANSS/NEOS) supply only methodology here and are not evaluated.