Searches for neutrinoless double-beta decay use germanium crystals to look for an event that deposits all of its energy at a single point. The main background comes from gamma rays that bounce around the crystal and deposit the same energy at two or more points. Current practice tells these apart with one number, the height of the current pulse divided by its energy. That number fails when the separate charge clouds reach the electrode close together in time, because their current spikes merge into what looks like one spike.
Speech processing has spent decades on a closely analogous problem: finding the exact sample at which each vocal-fold closure (an "epoch") excites the vocal tract, even when successive excitations overlap and ride on slow drifts. Those methods do not need a model of what each excitation's response looks like. This proposal transfers epoch-detection tools from speech to germanium current pulses. The goal is to count charge-arrival events and time them without a pulse template, and to test whether that beats both the scalar cut and learned models in the setting where templates are least trustworthy.
Core Claim
The claim is that template-free epoch estimators from speech can resolve two-site events at smaller drift-time separations than scalar current-amplitude cuts, and can transfer from simulation to real detector data better than methods that depend on simulated pulse shapes. The two speech estimators are zero-frequency filtering, which marks impulse locations as zero-crossings of a very-low-frequency filtered signal, and group-delay analysis, which reads an impulse's timing from the phase of a sharply decaying analysis window.
The second half of the claim matters most. A two-pulse template fit would be the statistically natural tool if the templates were right. In germanium detectors they are not: ideal drift simulation matches the real peak-current distribution at only 12.4% histogram overlap, and matches it at only 36.8% after learned translation, per Ad-hoc Pulse Shape Simulation using Cyclic Positional U-Net. A method that locates excitations without knowing their shape therefore has a principled reason to win, and that reason can be tested.
Where This Sits
The rare-event side supplies the problem. Multisite event discrimination for the majorana demonstrator establishes the current-amplitude-over-energy cut ("A/E"). It retains (90 ± 3.5)% of single-site events from the thallium-208 double-escape peak, cuts the single-escape peak to about 6%, and names spatial mismatch between calibration and signal events as its dominant uncertainty. Deep learning based pulse shape discrimination for germanium detectors compresses 256-sample waveforms into a 7-dimensional autoencoder code and matches A/E, at 90% background rejection and 90% signal acceptance. It also reports a class of multi-site events whose high current peaks fool A/E. That paper motivates the proposal, because it shows that information beyond the peak exists. It gives no physical account of that information, and its learned code has no units that can be audited.
The one existing use of time-frequency analysis in this pipeline, the wavelet features in Machine learning-powered data cleaning for LEGEND, targets electronic noise rather than event topology. It uses fixed dyadic scales, which lack sample-level timing.
The speech side supplies the method. Zero-frequency filtering, as used in Zero Frequency Filter Based Analysis of Voice Disorders, passes the differenced signal through two resonators at 0 Hz and then subtracts a local mean. Positive zero-crossings mark epochs, and the slope at each crossing gives an excitation strength. Cycle-level features from this filter reached 92.87% on voice-pathology classification against 66.66% for frame-averaged measures, and they held up under low-SNR noise. Zero-time windowing, as used in Identification and Classification of Fricatives in Speech Using Zero Time Windowing Method, applies a sharply decaying window at every sample and reads structure from the group delay. It identified 97.5% of sibilant boundaries, and its features scored 93.2% against 70% for standard cepstral features.
That same speech work also constrains the design. Zero-time windowing fails on non-sibilant fricatives, whose weak, diffuse spectra are mistaken for silence. The analogous risk here is that a low-energy secondary Compton site will vanish into noise. Zero-frequency filtering also needs a prior on the spacing between excitations to set its trend-removal window, and a germanium pulse has no pitch period. Both constraints shape the method below.
Research Program
The pipeline has three stages, all deterministic and all tuned only on calibration peaks.
The first stage adapts zero-frequency filtering to single transients. The trend-removal window is set to the measured width of the single-site current pulse, taken from double-escape-peak events for each detector. This replaces the speech method's pitch-period prior with a physically measured detector constant. Candidate charge-arrival epochs are positive zero-crossings, each with a strength given by its slope.
The second stage refines each candidate using the average group delay of a zero-time window placed on the current pulse. The window is $w[n]=1/(4\sin^2(\pi n/2N))$ for $n=1,\dots,N$, where $N$ is the window length in samples. The timing offset $\hat t$ is
$$\hat t = \frac{1}{\pi}\int_0^{\pi} \tau(\omega)\,d\omega, \qquad \tau(\omega) = -\frac{d}{d\omega}\arg X_w(\omega),$$
where $X_w$ is the Fourier transform of the windowed segment and $\tau(\omega)$ is its group delay. Stepping the window sample by sample, an impulse is located where $\hat t$ crosses zero. This gives sub-sample timing from phase rather than amplitude.
The third stage turns the epochs into a small set of physically labeled features. These are the number of epochs, the largest separation between them in nanoseconds, the ratio of secondary to primary strength, and the residual energy left after removing a single data-derived pulse at the primary epoch. The features feed either a two-dimensional cut or a classifier with fewer than 200 parameters, which keeps the model comparable in size to the autoencoder's classifier.
The analysis half of the program maps the method's resolution limit. In simulation, two-site events are generated with controlled separation in drift time and controlled energy ratio. For each method, the program measures the probability of detecting a second site across that grid, at the threshold that keeps 90% of single-site events. The resulting curves show exactly where each method stops seeing a second site. They also let the non-sibilant failure mode be quantified directly as a minimum detectable secondary energy.
Baselines
The do-nothing reference is the energy cut alone, with no pulse-shape discrimination. It fixes the multi-site survival that every method must improve on.
The strongest established method is the A/E cut of Multisite event discrimination for the majorana demonstrator, including its quadratic energy correction. It is reported at (90 ± 3.5)% double-escape-peak acceptance with about 6% single-escape-peak survival. It will be recomputed on our exact events, with the same energy correction fitted on the same calibration split. Beating it means lower single-escape-peak and full-energy-peak survival at matched 90% double-escape-peak acceptance, by more than the statistical uncertainty.
The strongest learned method is the autoencoder classifier of Deep learning based pulse shape discrimination for germanium detectors, reported at 90% full-energy-peak rejection with 90% double-escape-peak acceptance. It will be retrained on our waveforms with the same labeled calibration peaks and a hyperparameter budget equal to ours. We will report it both trained on real calibration data and trained on simulation then applied to real data.
The model-based competitor is a two-pulse template likelihood fit. It uses the drift simulator's templates, and a second version uses templates translated by CPU-Net.
Because the file reports its results only for noise cleaning, the wavelet feature set of Machine learning-powered data cleaning for LEGEND will be retrained as a discriminator. This isolates "any time-frequency features" from "epoch-synchronous features." No reported pulse-shape-discrimination result for it was found on Lacuna.
Two ablations isolate the key ingredient. The first replaces group-delay refinement with magnitude peak-picking on the same zero-frequency-filtered signal. The second replaces zero-frequency filtering with a smoothed derivative and a threshold. If either ablation matches the full method, the phase-based machinery is not what earns the result.
Why not just…?
Why not just keep A/E, plus delayed-charge recovery for alphas? For surface alphas, nothing needs fixing: delayed-charge recovery already rejects all of them beyond 3 mm with 0.2% bulk loss, per α-event characterization and rejection in point-contact HPGe detectors. This proposal therefore leaves alphas alone. For multi-site gammas, the autoencoder result shows that A/E misses events with high secondary peaks. The resolution curves will show whether those misses are closely spaced two-site events that epochs can recover. If the curves show no gain over A/E at any separation, the claim is false.
Why not just train a CNN or autoencoder on raw waveforms? On in-distribution real calibration data, a learned model may well match or beat us, and we will say so if it does. The case for the epoch method rests on transfer and auditability. Learned models trained on simulation inherit a peak-current mismatch that learned translation only raises to 36.8% overlap. The sim-trained versus real-trained comparison above tests whether template-free features degrade less. If they do not, the transfer argument fails.
Why not just take the derivative and count peaks, or fit two templates? Derivative peak-counting is the ablation. If it ties the full method, group delay adds nothing. The two-template fit is the real threat whenever its templates are accurate. We therefore run it with both simulated and translated templates. The proposal's thesis predicts that the fit wins in simulation but loses on real data. If the fit wins on real data too, the correct conclusion is that better templates, not speech methods, are the path forward.
Proposed Experiments
The simulation study uses drift-simulator pulses, following the 65,536-pulse single-site and multi-site setup of Ad-hoc Pulse Shape Simulation using Cyclic Positional U-Net, with added two-site grids. It produces resolution curves for every method. The real-data study uses the thorium-228 laboratory broad-energy germanium data from the same paper, with its 7:3 split, and seeks point-contact calibration data through collaboration. It reports single-escape-peak and full-energy-peak survival at 90% double-escape-peak acceptance, and ROC curves.
The first decisive check is the real-data survival comparison against A/E and the autoencoder at matched acceptance.
The second decisive check is the sim-to-real transfer gap. Every method is tuned on simulation and evaluated on real data, and the drop is compared.
The third check is stability. We ask whether single-site acceptance stays flat from 100 keV to 3 MeV without the ad hoc energy correction that A/E requires. We also ask whether acceptance agrees between double-escape-peak events and the two-neutrino double-beta continuum, a bulk single-site population where it is available. The second comparison probes the ±3.5% spatial systematic directly, because an epoch count should not depend on where a single site sits.
A null result on all three checks would itself settle a question the field currently answers only by assertion.
References
- Ad-hoc Pulse Shape Simulation using Cyclic Positional U-Net
- Multisite event discrimination for the majorana demonstrator
- Deep learning based pulse shape discrimination for germanium detectors
- Machine learning-powered data cleaning for LEGEND: a semi-supervised approach using affinity propagation and support vector machines
- Zero Frequency Filter Based Analysis of Voice Disorders
- Identification and Classification of Fricatives in Speech Using Zero Time Windowing Method
- $$\alpha $$-event characterization and rejection in point-contact HPGe detectors