Zero frequency filtering (ZFF) finds the moments when the vocal folds snap shut, called glottal closure instants or "epochs", by integrating speech so heavily that only energy near 0 Hz survives. The integration is unstable. Its output carries a large, slowly growing drift, and the method removes that drift by subtracting a moving average whose length must roughly equal one pitch period. That one tuning choice causes much of ZFF's fragility on singing, emotional speech and breathy or creaky voices, where the pitch period changes quickly or is poorly defined. Particle physicists face a structurally similar problem when they pull a narrow resonance out of a smooth "continuum" background. Their mature answer is not to smooth the background away. They fit a known smooth background shape and a signal model together, so that neither biases the other, and they calibrate how often background alone mimics a signal by using control data where the signal is known to be absent.
We propose to carry that practice over to ZFF. The core observation is that, inside a short analysis block, the output of the zero-frequency integrators is approximately a low-order polynomial (the accumulated drift from everything before the block) plus a new polynomial piece that begins at each glottal closure. That makes the ZFF output a spline whose knots sit at the epochs. Epoch detection then becomes knot detection in a joint background-plus-signal fit, and it needs no pitch-period window. Unvoiced and silent stretches serve as "off-resonance" control data that turn each detected knot into a calibrated significance, which addresses ZFF's spurious detections in noise.
Core Claim
Write the twice-integrated signal from standard ZFF as $y_2[n]$. Pre-differencing turns a glottal impulse at time $\tau$ into a two-sample pulse, and the four integrating poles at $z=1$ then respond to it with a function that grows quadratically in $(n-\tau)$ after $\tau$. Within a block of length $L$, this suggests the model
$$y_2[n] = \sum_{k=0}^{3} c_k\, T_k(n/L) + \sum_{j} a_j\,(n-\tau_j)_+^2 + e[n],$$
where $T_k$ are Chebyshev polynomials modelling the drift inherited from before the block, $(u)_+ = \max(u,0)$, $\tau_j$ are candidate epochs with strengths $a_j$, and $e[n]$ collects everything else. That residual includes formant leakage the integrators did not fully suppress and the departure of real glottal pulses from ideal impulses. We fit the drift coefficients $c_k$ and a sparse set of knot strengths $a_j$ jointly, by penalized least squares with a sparsity penalty on the $a_j$ over a dense grid of candidate $\tau_j$, followed by a refit on the selected knots.
The claim is that this joint fit locates epochs as accurately as tuned ZFF on modal speech. It should degrade much less when pitch varies or phonation is non-modal, because nothing in it assumes a period length. It also yields, as a by-product, an explicit drift estimate and a per-epoch strength $a_j$ that plays the role of ZFF's strength of excitation.
Relation to Prior Work
The motivating evidence comes from the ZFF literature itself. Evaluation of glottal epoch detection algorithms on different voice types compared six detectors on a seven-voice-quality dataset and found that ZFF suffers more misses and false alarms on breathy and lax voices than SEDREAMS. A Robust and Alternative Approach to Zero Frequency Filtering Method for Epoch Extraction identified the trend window as a failure point under fluctuating pitch. Its fix was to drop trend removal altogether, using a stable finite filter with many zeros at half the sampling rate. That filter still needs its zero count matched to the expected pitch period, so the period dependence moves into the filter design rather than disappearing. We keep the unstable integrators and instead model the drift explicitly. This also bears on the tension in that literature: Identification of F1 and F2 in Speech Using Modified Zero Frequency Filtering and Unsupervised Voice Activity Detection by Modeling Source and System Information using Zero Frequency Filtering show that the multi-scale detrended outputs carry useful system information. Their window choices, such as fractions of the pitch period, remain heuristic.
The closest existing method is Improved Epoch Extraction from Telephonic Speech Using Chebfun and Zero Frequency Filtering. It uses Chebyshev polynomials to represent the Hilbert envelope of band-limited speech before integration, but only as a numerical device, and it still detrends with a moving average afterwards. We use Chebyshev polynomials for a different job: as the explicit background model, fitted jointly with the epochs, in place of the moving average.
The method's template comes from collider analyses. Measurements of the branching fractions B(B−→Λ¯c−Ξc′0) , B(B−→Λ¯c−Ξc(2645)0) and B(B−→Λ¯c−Ξc(2790)0) extracts small signals from Belle data with unbinned likelihood fits in which a smooth background and the signal are modelled together. Branching-fraction measurements of $\psi(2S)$ decay to baryon-antibaryon final states and New measurements of $\Upsilon(1S)$ decays to charmonium final states estimate the background from data taken where the resonance is absent, scaled by exposure. We borrow both ideas: the joint fit, and an off-resonance control sample for calibration.
The physics literature also constrains the design. The interference problem it raises is that background and signal amplitudes add rather than their counts. For this reason we fit in the signal domain, $y_2[n]$, where the contributions are genuinely additive, and not on magnitude spectra. Spectral subtraction makes the magnitude-domain mistake, and Adaptive filtering for attenuating musical noise caused by spectral subtraction documents the artefacts that result. Physics also warns about peaking backgrounds that imitate signal. The speech analogue is a fricative onset or a noise burst, and that is exactly why we add the control-sample calibration.
Research Program
The work has three parts.
First, we build the joint fit. It runs on overlapping blocks of about 20–30 ms, longer than the lowest expected pitch period but short enough that four drift terms suffice. Candidate knots lie on a grid at the sampling interval. The sparsity weight is chosen once on held-out data and never per speaker. Knots in the overlap between blocks are merged.
Second, we add off-resonance calibration. Segments that a voice activity detector marks as unvoiced or silent go through the identical pipeline. The distribution of their fitted knot strengths, rescaled by segment duration in the same way physicists scale off-resonance yields by luminosity, gives an empirical null. Each knot found in voiced speech then receives a false-alarm probability, and a single significance threshold replaces ad hoc energy gating.
Third, we analyse where the knot model breaks. On the voice-quality data we measure how much the residual $e[n]$ grows for breathy and creaky phonation, where closure is gradual rather than impulsive. If needed, we replace the sharp $(u)_+^2$ basis with a smoothed version whose width is fitted to the data.
Baselines
The naive reference point is standard ZFF with a fixed trend window of 1–1.5 times the average pitch period, applied twice, as described in On Modeling Glottal Source Information for Phonation Assessment in Parkinson's Disease. On CMU Arctic, A Robust Non-Parametric and Filtering Based Approach for Glottal Closure Instant Detection reports ZFF at an identification rate of 96.62% with 3.29% false alarms. We rerun ZFF ourselves on the same utterances and EGG references, and we tune its window on the same held-out split used for our sparsity weight.
The strongest existing competitor on clean speech is the same paper's raised-cosine-filter method with histogram segregation. It reports an identification rate of 99.16% with 0.49% false alarms on CMU Arctic, and an identification rate still above 99% under −10 dB white noise. SEDREAMS is reported there at 98.44% and 1.15%, and we include it because Evaluation of glottal epoch detection algorithms on different voice types found it best overall across voice qualities. For telephone-band speech the strongest reference is Chebfun-ZFF. It reports an identification rate of 96.42% with a 0.29% miss rate for speaker SLT, and 69% identification accuracy within ±0.25 ms for speaker JMK, against 29.5% for Hilbert-envelope ZFF. The finite-filter variant of A Robust and Alternative Approach to Zero Frequency Filtering Method for Epoch Extraction is also a baseline; no reported identification numbers were found on Lacuna for it. All methods are scored against the same EGG-derived references with the same matching tolerance.
On clean CMU Arctic we do not expect to beat the raised-cosine method's reported figures, and we would count matching ZFF-plus-tuning as acceptable. Success means a clear reduction in misses and false alarms relative to both ZFF and SEDREAMS on the breathy, lax, creaky and falsetto subsets of the voice-quality data and on the emotional subsets of EMO-DB, together with no loss on modal speech. A matched gain on telephone-band speech over Chebfun-ZFF would be a bonus.
Two ablations isolate the key ingredients. In the first, we fit only the polynomial drift with no knot terms and then subtract it, which is the analogue of sideband subtraction without a signal model. This tests whether the joint fit matters at all. In the second, we remove the off-resonance calibration and use a fixed strength threshold, which isolates the false-alarm benefit.
Why not just…?
Why not just make the ZFF window adapt to a running pitch estimate? This is the most serious threat, and it is included as a baseline. The pitch-adaptive version first estimates the period with ZFF or an autocorrelation tracker, then sets the detrending window from that estimate. It inherits every failure of the pitch estimator, and those failures are concentrated where pitch is irregular or ill-defined: creaky voice, onsets and pitch jumps. Those are exactly the conditions where Evaluation of glottal epoch detection algorithms on different voice types saw ZFF degrade. If pitch-adaptive ZFF matches the joint fit on the creaky and emotional subsets, the window-free argument fails. We would then report the joint fit's value as limited to its calibrated significances.
Why not just high-pass or Savitzky–Golay-detrend the ZFF output? Any detrender that does not model the epochs treats the quadratic piece added at each closure as part of the trend. Part of the signal is therefore subtracted, and the zero-crossing shifts. This is the same bias that sideband interpolation under a peak suffers when the background model ignores the signal. The drift-only ablation is this alternative, and it directly tests whether modelling the knots changes timing accuracy.
Why not just use LPC residuals or sparse deconvolution of the raw speech? Inverse filtering needs a vocal tract model. It fails when the fundamental frequency approaches the first formant, as noted in Estimation of Fundamental Frequency from Singing Voice Using Harmonics of Impulse-like Excitation Source, and residual-peak detectors such as DYPSA degrade on breathy voices. Our sparse fit runs after the zero-frequency integrators have already suppressed the formants, so it needs no vocal tract model. A reviewer may fairly call the method sparse deconvolution in the integrated domain. We accept that description. The empirical question is whether that domain is better, which the singing and high-pitch comparisons will answer.
Why not just drop the trend with the finite-filter ZFF? That filter's zero count is tied to the expected pitch period, and it spreads in time when lengthened. It also discards the drift, which the voice activity and formant work shows is informative. It is run as a baseline under identical scoring.
Proposed Experiments
The primary test uses CMU Arctic speakers with EGG ground truth, measuring identification rate, miss rate, false alarms and timing accuracy within ±0.25 ms. We repeat it under added white and babble noise from 20 dB down to −10 dB, and under the 8 kHz, 300–3400 Hz telephone simulation from the Chebfun-ZFF paper.
The decisive robustness check uses the male voice-quality dataset with its seven phonation types and synchronized EGG, together with EMO-DB under 10 and 20 dB noise. The comparison against pitch-adaptive ZFF on the creaky and high-arousal emotional material is the single result that most determines whether the idea holds.
A calibration check uses noise-only and unvoiced segments from CMU Arctic held out from the null construction. For the claim of calibration to stand, the empirical false-alarm rate at each nominal significance level must track that nominal level.
Finally, we run a sensitivity sweep over block length and polynomial order. If results depend on block length about as strongly as ZFF depends on its trend window, the method has only relocated the tuning problem, and we will say so.
References
- Evaluation of glottal epoch detection algorithms on different voice types
- A Robust and Alternative Approach to Zero Frequency Filtering Method for Epoch Extraction
- Identification of F1 and F2 in Speech Using Modified Zero Frequency Filtering
- Unsupervised Voice Activity Detection by Modeling Source and System Information using Zero Frequency Filtering
- Improved Epoch Extraction from Telephonic Speech Using Chebfun and Zero Frequency Filtering
- Measurements of the branching fractions B(B−→Λ¯c−Ξc′0) , B(B−→Λ¯c−Ξc(2645)0) and B(B−→Λ¯c−Ξc(2790)0)
- Branching-fraction measurements of<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi>ψ</mml:mi><mml:mo stretchy="false">(</mml:mo><mml:mn>2</mml:mn><mml:mi>S</mml:mi><mml:mo stretchy="false">)</mml:mo></mml:math>decay to baryon-antibaryon final states
- New measurements of<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi>Υ</mml:mi><mml:mo stretchy="false">(</mml:mo><mml:mn>1</mml:mn><mml:mi>S</mml:mi><mml:mo stretchy="false">)</mml:mo></mml:math>decays to charmonium final states
- Adaptive filtering for attenuating musical noise caused by spectral subtraction
- On Modeling Glottal Source Information for Phonation Assessment in Parkinson’s Disease
- A Robust Non-Parametric and Filtering Based Approach for Glottal Closure Instant Detection
- Estimation of Fundamental Frequency from Singing Voice Using Harmonics of Impulse-like Excitation Source