PaperlayerView exact revision on arXiv

Source-linked AI summary

Detection of Glottal Closure Instants from Speech Signals: a Quantitative Review

Thomas Drugman, Mark Thomas, Jon Gudnason, Patrick Naylor, Thierry Dutoit

arXiv:2001.00473v1cs.SDcs.CLeess.AS

TL;DR

Precise GCI locations are needed to exploit voiced-speech pseudoperiodicity in speech processing. The paper compares five waveform-based detectors against EGG references across six databases and tests clean, noisy, reverberant, and deconvolution settings; SEDREAMS and YAGA lead on clean speech, while ZFR and SEDREAMS are most robust to degradation.

  • Problem

    The paper addresses the need for reliable and accurate automatic GCI detection from speech waveforms across clean, noisy, and reverberant conditions.

  • Method

    The authors compare HE, ZFR, DYPSA, SEDREAMS, and YAGA against contemporaneous EGG recordings across six speech databases and a causal-anticausal deconvolution application.

  • Results

    SEDREAMS and YAGA perform best on clean speech, exceeding 98% identification on every database, while ZFR and SEDREAMS show the highest robustness to noise and reverberation.

  • Takeaways & Limitations

    For unknown environments, the authors suggest SEDREAMS because it performs strongly on clean speech, degraded conditions, reverberation, and real-time implementation.

Abstract

from arXiv · show

The pseudo-periodicity of voiced speech can be exploited in several speech processing applications. This requires however that the precise locations of the Glottal Closure Instants (GCIs) are available. The focus of this paper is the evaluation of automatic methods for the detection of GCIs directly from the speech waveform. Five state-of-the-art GCI detection algorithms are compared using six different databases with contemporaneous electroglottographic recordings as ground truth, and containing many hours of speech by multiple speakers. The five techniques compared are the Hilbert Envelope-based detection (HE), the Zero Frequency Resonator-based method (ZFR), the Dynamic Programming Phase Slope Algorithm (DYPSA), the Speech Event Detection using the Residual Excitation And a Mean-based Signal (SEDREAMS) and the Yet Another GCI Algorithm (YAGA). The efficacy of these methods is first evaluated on clean speech, both in terms of reliabililty and accuracy. Their robustness to additive noise and to reverberation is also assessed. A further contribution of the paper is the evaluation of their performance on a concrete application of speech processing: the causal-anticausal decomposition of speech. It is shown that for clean speech, SEDREAMS and YAGA are the best performing techniques, both in terms of identification rate and accuracy. ZFR and SEDREAMS also show a superior robustness to additive noise and reverberation.

I. INTRODUCTION

Glottal-synchronous processing exploits voiced-speech pseudoperiodicity, but requires reliable GCI detection across clean and degraded speech. This paper reviews five methods, evaluates them against EGG references, and examines a practical deconvolution application.

  • Motivation: GCI locations enable pitch-contour estimation and delineation of individual speech cycles for several speech-processing applications.Applications include prosodic modification, dereverberation, glottal-flow estimation, synthesis, voice-source modelling, and causal-anticausal deconvolution.
  • Motivation: Automatic GCI detection is needed for clean speech and speech corrupted by acoustic noise or reverberation.Earlier autocorrelation methods were unreliable because formant frequencies caused multiple maxima.
  • Methodological context: Smoothing attenuates vocal-tract resonances, additive noise, and reverberation while preserving periodicity, but can make precise GCI timing ambiguous.LP residual information can supplement smoothed speech to improve timing estimates.
  • Scope and contribution: The paper objectively evaluates HE, DYPSA, ZFR, SEDREAMS, and YAGA against contemporaneous EGG-based reference GCIs from six databases.The databases contain 232 minutes of speech and EGG recordings, with additional tests under additive noise, reverberation, and causal-anticausal deconvolution.
  • Methodological context: The HE method computes a Hilbert envelope from the LP residual and uses a Center of Gravity signal whose negative zero-crossings indicate GCIs.The CoG-based signal is intended to remove ambiguity from discontinuities around GOIs.

B. The DYPSA algorithm

DYPSA detects candidate GCIs from the group delay of the LP residual and refines them using phase-slope information before dynamic-programming selection.

  • Candidate detection: DYPSA combines LP-residual group-delay candidate detection with N-best dynamic programming.The LP residual is used to identify impulsive features associated with GCIs.
  • Group Delay Function: The group delay is the average slope of the unwrapped short-time Fourier-transform phase spectrum of a windowed LP-residual segment.Negative-going zero crossings select GCI candidates.
  • Group Delay Function: Energy-weighted group delay provides an efficient time-domain formulation that behaves like a bounded center of gravity of the windowed residual segment.Its negative-going zero crossings estimate peak locations.
  • Phase-slope projection: When impulsive features do not produce negative-going zero crossings, DYPSA estimates their locations from the midpoint between nearby local minima and maxima.This phase-slope projection addresses missing zero crossings in the group-delay signal.

2) Dynamic Programming:

DYPSA’s dynamic-programming stage selects a plausible sequence of GCI candidates by minimizing a weighted cost function based on speech and pitch characteristics.

  • Dynamic Programming: Erroneous GCI candidates are removed by selecting the candidate subset that minimizes a dynamic-programming cost function.The selected subset is intended to correspond most closely to true GCIs.
  • Cost elements: Waveform-similarity cost penalizes candidates uncorrelated with the previous candidate.It encourages continuity between neighboring candidates.
  • Cost elements: Pitch-deviation cost penalizes large deviations between the current and previous two candidates.This encodes expected pitch continuity.
  • Cost elements: Projected-candidate cost penalizes candidates arising from potentially erroneous phase-slope peaks.The cost targets candidates introduced by phase-slope projection.
  • Cost elements: Normalized-energy cost penalizes candidates that do not coincide with high speech-signal energy.Ideal phase-slope deviation instead favors candidates from zero crossings with gradients close to unity.

C. The Zero Frequency Resonator-based technique

ZFR detects GCIs by filtering speech through cascaded zero-frequency resonators, removing the resulting trend, and exploiting the filtered signal’s pitch-related oscillation. SEDREAMS instead narrows and refines expected GCI intervals.

  • ZFR technique: ZFR exploits the broadband impulsive nature of GCI excitation by analyzing speech through zero-frequency resonators, minimizing vocal-tract-resonance influence.The resulting signal is intended to primarily reflect the excitation.
  • ZFR technique: ZFR first removes recording-related DC or low-frequency bias, then passes the signal twice through an ideal zero-frequency resonator.The two passes reduce vocal-tract-resonance influence in the output.
  • ZFR technique: Repeated mean subtraction removes the exponentially increasing or decreasing trend from the twice-filtered signal.Mean removal is repeated three times to avoid residual drift.
  • ZFR technique: A mean-subtraction window of about 1 to 2 average pitch periods is considered adequate, and this study uses 1.5 average pitch periods.The window length was reported as not very critical within that range.
  • SEDREAMS technique: SEDREAMS locates GCIs in two stages: identifying short expected-GCI intervals and refining locations within those intervals.The present study omits its GOI-location procedure because it focuses only on GCIs.

1) Determining intervals of presence using a mean-based signal:

SEDREAMS uses a mean-based signal to identify pitch-synchronous intervals where GCIs are likely to occur, then bounds their timing uncertainty. Because GCI position varies within the signal cycle, the intervals provide reliable candidate regions rather than exact locations.

  • The mean-based signal y(n) is obtained by applying a windowed mean operation to the speech waveform s(n).
  • The signal oscillates at the local pitch period, but inappropriate window lengths can create false alarms or misses.
  • Actual GCIs may occur at non-constant relative positions within a normalized mean-based cycle across speakers.
  • Presence intervals begin at a mean-based-signal minimum and extend for 0.35 times the local pitch period.

2) Refining GCI locations using the residual excitation:

SEDREAMS refines broad presence intervals by locating the largest LP-residual discontinuity, while YAGA derives candidates from a voice-source estimate and selects among them using dynamic programming.

  • Refining GCI locations using the residual excitation: SEDREAMS inspects the LP residual within each presence interval and assigns the GCI to its largest discontinuity.
  • Refining GCI locations using the residual excitation: Combining mean-based-signal intervals with LP-residual peak picking enables accurate and unambiguous GCI detection in the illustrated example.
  • Refining GCI locations using the residual excitation: The mean-based intervals bound SEDREAMS timing error by 0.35 times the local pitch period, while also limiting misses and false alarms.
  • The YAGA algorithm: YAGA uses N-best dynamic programming to select a path through GCI candidates derived from an LP-based voice-source estimate.
  • The YAGA algorithm: YAGA applies stationary-wavelet multiscale analysis to the voice-source signal, producing impulse-like discontinuity features while reducing vocal-tract-resonance effects.
  • The YAGA algorithm: Negative-going group-delay zero-crossings form YAGA’s candidate set, after which dynamic programming distinguishes GCIs, GOIs, and false candidates.

III. ASSESSMENT OF GCI EXTRACTION TECHNIQUES

The evaluation compares GCI detectors against synchronized EGG references across six speech corpora and also tests their value for causal-anticausal decomposition. Reliability and timing accuracy are assessed with complementary measures.

  • A. Speech Material: Six speech corpora with contemporaneous EGG recordings provide ground truth for comparing the detection methods.
  • A. Speech Material: The databases include single-speaker CMU ARCTIC, nonsense-word, KED Timit, and multi-speaker APLAWD recordings sampled at 16 kHz.
  • Comparison with Electroglottographic Signals: The evaluation compares estimated GCI locations with synchronized EGG-derived reference positions and additionally tests causal-anticausal deconvolution.
  • Comparison with Electroglottographic Signals: Identification Rate measures cycles with exactly one detected GCI, while Miss Rate and False Alarm Rate measure zero or multiple detections.
  • Comparison with Electroglottographic Signals: Identification Accuracy is the timing-error standard deviation, and Accuracy to ± 0.25 ms is the proportion of detections within that error bound.

2) A Speech Processing Application: the Causal-Anticausal Deconvolution:

Causal-anticausal decomposition separates speech into minimum- and maximum-phase components, with the anticausal component representing the glottal contribution. Its sensitivity to GCI errors makes it an application-level test of GCI extraction.

  • Causal-anticausal decomposition is a non-parametric source-tract deconvolution method that is highly sensitive to GCI location errors.
  • The decomposition separates minimum-phase and maximum-phase components, with the maximum-phase component attributed to the glottal contribution.
  • Isolating the anticausal component enables estimation of the glottal open phase.
  • The ZZT and CCD algorithms are functionally equivalent, while CCD was recommended for its much higher computational speed than ZZT.
  • The spectral center of gravity of the anticausal component provides a criterion for deciding whether a frame was correctly decomposed.
  • The application is expected to show fewer incorrectly decomposed frames when supplied with good GCI estimates.

IV. EXPERIMENTS ON CLEAN SPEECH DATA

On clean speech, SEDREAMS and YAGA generally provide the strongest GCI detection reliability and accuracy across the evaluated databases. Their estimates also perform well in causal-anticausal deconvolution, although precise synchronization alone is insufficient for correct separation.

  • More than 98% of GCIs are correctly identified by SEDREAMS on every dataset, while YAGA reaches 95.70% on the RAB database.
  • YAGA leads the accuracy measure on all databases except RAB for frames whose timing error is below 0.25 ms.
  • SEDREAMS has nearly comparable accuracy to YAGA, while DYPSA and HE are outperformed by both methods across all datasets.
  • SEDREAMS and YAGA produce the sharpest and highly similar distributions of GCI timing error averaged across datasets.
  • SEDREAMS and YAGA generally outperform the other GCI detection methods on clean speech in both reliability and accuracy.
  • In mixed-phase deconvolution, SEDREAMS and YAGA outperform other approaches on all datasets except SLT, but precise GCI synchronization is necessary rather than sufficient for correct separation.

V. ROBUSTNESS OF GCI EXTRACTION METHODS

The robustness experiments evaluate GCI detection under additive noise and reverberation-related real-world degradation. Under additive noise, ZFR and SEDREAMS are the most robust overall, while DYPSA degrades most strongly and HE has the worst identification accuracy.

  • Additive noise was tested at various SNRs using both white Gaussian noise and babble noise, with results averaged over the six databases.
  • White noise causes more severe reliability degradation, whereas babble noise affects accuracy more strongly.
  • SEDREAMS and ZFR provide the best reliability robustness, with performance almost unchanged down to 0 dB SNR.
  • ZFR is the most robust for accuracy, while DYPSA shows the strongest degradation and HE the worst identification accuracy.
  • The robustness of ZFR and SEDREAMS is attributed to the low noise sensitivity of zero-frequency resonators and the mean-based signal.
  • DYPSA and YAGA are strongly sensitive to additive noise, plausibly because their thresholds were optimized for clean speech.

B. Robustness to Reverberation

Reverberation progressively degrades GCI detection, with the largest performance change between T60 = 100 and 200 ms. ZFR and SEDREAMS are comparatively robust, while no single algorithm performs consistently well across all measures.

  • Performance monotonically decreases as reverberation increases, with the largest change occurring between T60 = 100 and 200 ms.
  • Reverberation particularly harms identification rates for the LP-based HE, DYPSA, and YAGA approaches.Spurious residual peaks generally increase false alarms for DYPSA and YAGA, but increase miss rates for HE.
  • ZFR is least sensitive to reverberation and provides the best overall performance, while SEDREAMS generally offers better identification rates than other methods.
  • Although spurious peaks increase false alarms, the timing accuracy of correctly detected GCIs is much less affected.
  • Single-channel reverberant GCI detection remains challenging because no algorithm consistently performs well across all five measures.
  • Computational complexity was also compared using averaged Relative Computation Time across male and female speakers and all databases.Non-accelerated DYPSA was fastest at around 20% RCT, while ZFR was slowest at 75%; accelerated SEDREAMS reached about 6%.

VII. CONCLUSION

The paper compares five GCI detection methods across clean, noisy, and reverberant speech, including their use in causal-anticausal deconvolution and computational complexity. SEDREAMS and YAGA perform best on clean speech, while ZFR and SEDREAMS are most robust, motivating SEDREAMS for unknown environments.

  • Five GCI detection methods were evaluated on six databases containing approximately four hours of speech from male and female speakers.
  • On clean speech, SEDREAMS and YAGA achieved comparable best performance, exceeding 98% identification rates and over 80% accuracy within 0.25 ms for every database.
  • High-performing GCI detection significantly improved the proportion of correctly deconvolved frames in causal-anticausal decomposition.
  • ZFR and SEDREAMS showed the highest robustness to additive noise and reverberation, while reverberation dramatically affected detection precision.
  • For applications in unknown environments, the authors suggest SEDREAMS because of its clean-speech performance, robustness, and suitability for real-time implementation.
Loading 2001.00473v1…