Source-linked AI summary
Glottal Closure and Opening Instant Detection from Speech Signals
Thomas Drugman, Thierry Dutoit
TL;DR
The paper addresses the challenge of detecting GCIs and the less studied GOIs directly from speech waveforms. It proposes mean-based event intervals followed by LP-residual refinement, and reports improved GCI detection efficiency and noise robustness over DYPSA, with GOI identification also supported.
Problem
GOI detection has limited prior work because GOI excitation is weaker and more dispersed than GCI excitation, while GCI locations support several speech-processing applications.
Method
The procedure extracts expected GCI and GOI intervals from a mean-based signal and refines each event location using the LP residual.
Results
The method significantly improves GCI detection efficiency and noise robustness over DYPSA on the CMU ARCTIC database, while also determining GOIs.
Takeaways & Limitations
The procedure provides a waveform-based approach for locating both GCIs and GOIs, with promising GOI identification accuracy for glottal source characterization.
Abstract
from arXiv · showhide
This paper proposes a new procedure to detect Glottal Closure and Opening Instants (GCIs and GOIs) directly from speech waveforms. The procedure is divided into two successive steps. First a mean-based signal is computed, and intervals where speech events are expected to occur are extracted from it. Secondly, at each interval a precise position of the speech event is assigned by locating a discontinuity in the Linear Prediction residual. The proposed method is compared to the DYPSA algorithm on the CMU ARCTIC database. A significant improvement as well as a better noise robustness are reported. Besides, results of GOI identification accuracy are promising for the glottal source characterization.
1. Introduction
GCIs are important speech-processing events used in analysis and synthesis, while GOI detection remains challenging because opening excitations are weaker and more dispersed. The paper proposes a two-step waveform procedure to detect both events.
- Motivation: GCIs mark significant vocal-tract excitation during voiced speech and are important for speech processing.They correspond to moments of high energy in the glottal signal.
- Motivation: GCI locations are required by closed-phase LP analysis and assumed by applications including speech synthesis, voice transformation, and speaker identification.TD-PSOLA uses GCIs to help eliminate concatenation discontinuities or phase mismatches.
- Prior work: Existing GCI methods use diverse signals and criteria, including autocovariance determinants, LP-residual peaks, Frobenius norms, nonlinear prediction, and wavelets.Comparative studies identified DYPSA and another approach as outperforming other state-of-the-art methods.
- Motivation: GOI detection has received little attention because excitation energy at GOIs is weaker and more dispersed than at GCIs.A wavelet-product method was proposed, but it reported no quantitative results.
- Contribution: The proposed procedure first derives expected-event intervals from a mean-based signal, then refines locations using the largest LP-residual value within each interval.The LP residual peak is assumed to correspond to the strongest excitation impulse.
2. Proposed method
The method uses a mean-based signal to identify short candidate intervals for GCIs and GOIs, then uses the LP residual to assign accurate event locations. Interval boundaries are defined from extrema and zero-crossings, with an added margin for GOIs.
- Overview: The procedure has two successive steps: mean-based interval determination followed by LP-residual location refinement.The first step finds short regions for GCIs and GOIs; the second refines their accurate locations.
- Interval determination: The mean-based signal is computed from sliding-window speech segments to identify where speech events are expected.The approach is motivated by the claim that excitation discontinuities extend across the spectrum, including zero frequency.
- Interval determination: The windowing function has length 2N + 1, and the experiments use a Blackman window whose length is selected empirically.The window length is discussed in Section 3.1.
- Interval determination: GCI intervals span the mean-based signal’s minimum to the following positive zero-crossing, whereas GOI intervals span its maximum to the following negative zero-crossing.A 0.25 ms margin is added on both sides of the GOI interval, and preceding zero-crossings may be replaced by extrema midpoints.
- Location refinement: Within each candidate interval, the strongest LP-residual peak supplies the accurate GCI or GOI location.The LP residual is used because removing an estimated vocal-tract response is expected to expose significant excitation impulses.
3. Results
On the CMU ARCTIC database, the proposed method was evaluated for identification efficiency, timing accuracy, parameter selection, and noise robustness against DYPSA. It showed stronger overall performance, while GOI timing remained less precise than GCI timing.
- Experimental setup: The experiments used 16 kHz speech from three CMU ARCTIC speakers, with aligned EGG signals as reference and DYPSA as the comparison method.The database contained 1,132 phonetically balanced utterances per speaker, totaling approximately 2 hours 40 minutes.
- Evaluation measures: The evaluation measured Identification Rate, Miss Rate, False Alarm Rate, Identification Accuracy, and accuracy within ± 0.25 ms.Identification Accuracy was defined as the standard deviation of timing-error distributions.
- Window-length impact: A window length between 1.5 and 2 average pitch periods minimized misidentification, so the experiments used 1.75·T0,mean.Short windows produced spurious extrema and false alarms, whereas large windows increased misses through smoothing.
- Identification performance: The proposed method outperformed DYPSA across identification rates and timing-accuracy measures, except for speaker JMK, whose results were nearly similar.The identification comparison covered GCI and GOI detection, while timing accuracy differed between the two event types.
- Identification performance: 84% of identified GOIs had an absolute timing error below 1 ms, although GOIs were located less precisely than GCIs.The reported GOI result was considered efficient despite the weaker and more dispersed excitation at opening instants.
- Noise robustness: With white and babble noise, the proposed method remained almost insensitive to degradation up to 0 dB, whereas DYPSA began degrading from 30 dB and was severely affected from 10 dB.The same robustness pattern held for both noise types.
4. Conclusion
The paper introduced a two-step waveform-based procedure for detecting GCIs and GOIs. Experiments reported improved GCI efficiency and noise robustness over DYPSA, while GOI localization was promising but less precise than GCI localization.
- Conclusion: The procedure first extracts event intervals from a mean-based signal, then refines event locations using the LP residual.The first step supports identification, while the second supports timing accuracy.
- Conclusion: Compared with DYPSA on CMU ARCTIC, the method achieved significant improvement in GCI detection efficiency and noise robustness.The conclusion reports these improvements at the paper level without specifying individual metric values.
- Conclusion: GOI locations were obtained with encouraging precision, although they were not yet comparable to the precision achieved for GCIs.Future work targets improved GOI localization through open-quotient trajectories.
- Future work: The authors plan to combine the procedure with source-filter deconvolution methods to characterize the glottal source.This is identified as a future research direction rather than a demonstrated result.