Stable Excitation-Attenuated Linear Prediction, from Vowels to Scintillator Pulses

Many physical signals share one shape: a sharp kick followed by a free ring-down. A voiced vowel is a train of glottal pulses, and each pulse makes the vocal tract ring at its resonances. A liquid-scintillator neutrino detector records a burst of light when a particle deposits energy, and the tail then decays at rates set by the scintillator chemistry and by the type of particle. In both cases the useful physics sits in the ring-down, while the kick corrupts naive estimates. Speech researchers built a family of estimators for this situation. Weighted linear prediction fits an all-pole filter, meaning a model that predicts each sample from a few previous ones, while down-weighting samples near the kick. These estimators work very well, but they can return filters that are unstable (they grow instead of decay). One separate estimator solves the stability problem but does not use the kick-attenuating weights. We propose to join the two and then carry the result to a regime where stability is not just a numerical convenience but a physical law: the poles of a scintillation pulse must correspond to positive decay times.

The payoff on the physics side is indirect but real. Reactor-antineutrino experiments such as SNO+ measure the solar mass splitting $\Delta m^2_{21}$ from a few dozen inverse-beta-decay candidates. Their ability to reject backgrounds such as $(\alpha,n)$ reactions depends on pulse-shape discrimination: telling particle types apart from the time profile of the light (Allega et al. (2025)). A pulse-shape feature that stays reliable when the prompt peak is saturated, channels fail, or timing jitters would tighten background estimates in exactly the statistics-starved analyses where each event counts.

Core Claim

A sampled sum of $K$ decaying exponentials satisfies an exact order-$K$ linear recurrence. Linear prediction on a pulse tail therefore recovers its decay constants. This is Prony's classical idea, and we do not claim it as new. What is new is the estimator. It minimizes an excitation-attenuating weighted prediction error,

$$E = \sum_n w_n \Big(x_n - \sum_{k=1}^{p} a_k x_{n-k}\Big)^2 ,$$

where $x_n$ is the digitized waveform, $a_k$ are the $p$ prediction coefficients, and $w_n$ is small in a window after the excitation instant and large on the free decay. It also enforces, by construction, that every root $z_k$ of the predictor lies inside the unit circle. For a scintillator this constraint means

$$z_k = e^{-\Delta t/\tau_k}, \qquad |z_k|<1 \iff \tau_k>0 ,$$

with $\Delta t$ the sampling interval and $\tau_k$ a decay time. We claim three things. First, the stabilization mechanism of Stabilized Weighted Linear Prediction (SWLP) can carry pitch-synchronous attenuation weights with no loss of formant accuracy. Second, the resulting estimator is more robust than hard closed-phase masking to misplaced excitation instants. Third, its pole features give pulse-shape discrimination that degrades more gracefully than tail-to-total charge ratios under prompt-region distortions.

Where This Sits in Prior Work

Two strands of speech work supply the method. Magi et al. (2007) introduced SWLP. It modifies the matrix structure of weighted linear prediction so that the fitted filter is always stable, but it uses only short-time-energy weights. It fights the pull of pitch harmonics indirectly, through smoothing, and that smoothing broadens peaks (Gowda et al. (2013) had to add a group-delay step to re-sharpen them). Alku et al. (2012) introduced attenuated-main-excitation weighting (WLP-AME), which down-weights samples just after each glottal closure. Airaksinen et al. (2013) formalized the same idea as quasi-closed-phase analysis, and Gowda and Alku (2016) extended it to time-varying coefficients. These papers motivate our work: they show that attenuation weighting sharply reduces harmonic bias. They also constrain it. All of them lack a stability guarantee, and all depend on accurately located glottal closures. The research file finds no work that unifies SWLP stabilization with attenuation weights, and none that applies either technique to neutrino detectors.

On the physics side, the relevant tools are global fits and global regularizers. Daya Bay's Wiener-SVD unfolding (An et al. (2021)) and its covariance $\chi^2$ oscillation fits (An et al. (2023)) operate on energy spectra. SNO+ uses charge-integration pulse-shape discrimination (Allega et al. (2025)).

We deliberately reject one tempting idea: using attenuation weights to "suppress the 5 MeV bump" in reactor spectra. The bump is real reactor physics that scales with thermal power (An et al. (2016)). Down-weighting it is equivalent to dropping bins or adding a free-shape nuisance, both of which are trivial. Daya Bay's near-far design already cancels most of its effect on $\theta_{13}$. The waveform level is where the kick-then-ring-down analogy is exact.

Research Program

The first stage is algorithmic. We generalize SWLP's stabilizing construction from energy weights to an arbitrary non-negative weight profile. We then prove, or refute with a counterexample, that the construction's stability guarantee carries over to attenuation-shaped profiles, and that it does so even when some weights approach zero.

We also design a "soft-epoch" variant. Here the attenuation window is a smooth function of an estimated excitation time that carries an uncertainty, rather than a hard window at a detected instant. This directly addresses the weakness that misplaced closures destroy WLP-AME.

The second stage validates the estimator where ground truth exists: on physically simulated vowels, and on the Vocal Tract Resonance database of reference formant tracks. Speech is used here as a calibrated testbed, not as the destination.

The third stage transfers the estimator to scintillator waveforms. Because the file lists no public raw-waveform release, we will build a waveform simulator. It will draw photon arrival times from multi-exponential profiles, with parameters scanned around published scintillator values rather than fixed to numbers we do not have. The simulator will then add PMT single-photoelectron response, timing jitter, prompt saturation, random channel dropout mimicking the failed-PMT conditions that produced Daya Bay's muon-x background (Daya Bay Collaboration (2022)), and Poisson counting noise. In the tail, the weights also absorb inverse-variance factors.

As an optional fourth tier, a benchtop scintillator cell exposed to alpha and gamma sources would provide real waveforms. Finally, we propagate the measured change in background rejection into a toy SNO+-like $\Delta m^2_{21}$ fit. This measures how much the waveform-level gain matters for oscillation precision.

Baselines

The naive reference in speech is unweighted autocorrelation linear prediction. It is stable but suffers harmonic bias; Alku et al. (2012) report errors above 10% for high-pitched vowels. The strongest speech competitor is WLP-AME itself. For vowel [a] it reports relative formant errors of 3.0% (F1), 1.7% (F2), and 0.5% (F3), against 10.5%, 6.4%, and 1.9% for robust LP, all on physically synthesized vowels (Alku et al. (2012)). On real speech, the strongest reported number is from time-varying quasi-closed-phase analysis with an $L_1$ penalty: a 91.0% F3 detection rate with 5.7% F3 estimation error on the Vocal Tract Resonance database (Gowda and Alku (2016)). We will reimplement all speech methods with the same prediction order, sampling rate, and excitation instants, and give each the same grid search over window parameters. On the synthetic vowels, beating WLP-AME means matching its formant errors within resampling uncertainty while producing zero unstable filters. Under jittered excitation instants, it means lower errors than WLP-AME.

Two ablations isolate the key ingredient. The first is WLP-AME followed by post-hoc reflection of unstable poles. It tests whether built-in stability matters beyond simply patching the output. The second is SWLP with energy weights (Magi et al. (2007)). It tests whether attenuation weights add anything beyond stabilized smoothing.

For pulse shapes, the naive reference is the tail-to-total charge ratio used in SNO+-style analyses (Allega et al. (2025)). No reported result was found on Lacuna for pulse-shape discrimination figures of merit, so all numbers come from our own simulation. The strongest competitor is a template likelihood ratio that knows the true decay constants, including a nuisance parameter for saturation. Beyond that, we include unweighted Prony/linear prediction, hard-masked tail-only linear prediction, and a small convolutional classifier trained on the same simulated events. All methods see identical waveforms and receive equal tuning budgets on a held-out calibration split. Beating the field means a higher background rejection at fixed signal efficiency under distortions the methods were not tuned on, while staying within a few percent of the template on clean pulses.

Why not just…?

Why not just drop the samples near the excitation (closed-phase covariance analysis)? In speech, this approach fails at high pitch because the closed phase holds too few samples for a well-conditioned solve. That failure is the reason Alku et al. (2012) moved to soft weights. In scintillators the tail is long, so hard masking is a genuine threat. However, the samples right after the peak carry most of the information about the fast decay component, and discarding them costs discrimination power. The hard-mask baseline settles this directly. If it matches our estimator, we will report that the soft weights only matter for speech.

Why not just fit multi-exponential templates by nonlinear least squares or likelihood? With known shapes and clean data, this is close to optimal, and we expect it to win on undistorted pulses. Its weaknesses are non-convexity, sensitivity to initialization, and reliance on a shape model that aging hardware violates. The distortion-injection experiments test whether our closed-form estimator is more robust. If it is not, the remaining contribution is speed and guaranteed physical decay times, and we will say so.

Why not just train a neural classifier on waveforms? A classifier trained on simulation can overfit simulator details and gives no interpretable decay times to monitor over years of running. We test it on out-of-distribution distortions (dropout rates and saturation levels not seen in training). If it holds up there, it is the better background rejector, and our method's value narrows to interpretability and calibration monitoring.

Why not just reflect unstable poles after WLP-AME? Reflecting poles changes the spectral envelope (Magi et al. (2007)). In scintillators it silently turns a nonsensical growing mode into an arbitrary decay time. The reflection ablation measures both the frequency of instability and how much reflection distorts the output.

Decisive Checks

The first check is on the synthetic vowel sets used by Alku et al. (2012) and Airaksinen et al. (2013). There we count how often unconstrained WLP-AME returns unstable filters, and we verify that the stabilized version keeps formant errors at WLP-AME levels. If stabilization costs more than a small fraction of that accuracy, the unification fails.

The second check shifts excitation instants by controlled amounts and plots formant error against shift. The soft-epoch variant should degrade more slowly than both WLP-AME and hard masking.

The third check runs the scintillator simulator with increasing saturation and channel dropout. It compares background rejection at fixed signal efficiency across all baselines, with the template likelihood as the ceiling. It also verifies that recovered decay times stay unbiased.

The fourth check feeds each method's rejection curve into a toy reactor fit with SNO+-like statistics. It reports the change in the $\Delta m^2_{21}$ interval relative to the published $8.85^{+1.10}_{-1.33}\times10^{-5}$ eV² (Allega et al. (2025)). The proposal succeeds if the method is stable by construction, matches attenuation weighting on speech, and shows a measurable robustness margin over charge-ratio discrimination under realistic detector degradation.