Source-linked AI summary
A Comparative Study of Glottal Source Estimation Techniques
Thomas Drugman, Baris Bozkurt, Thierry Dutoit
TL;DR
Estimating glottal contribution from speech alone is a blind-separation problem, and natural speech provides no reference for quantitative accuracy. The paper compares three estimation approaches using synthetic speech and a large real-speech corpus. Mixed-phase decomposition and closed-phase inverse filtering perform best on clean synthetic and real speech, while iterative and adaptive inverse filtering is most robust in noisy conditions.
Problem
Glottal estimation from speech alone is a blind-separation problem, and natural speech has no target reference signal for quantitative assessment.
Source: §1 · p. 2Method
The study compares closed-phase inverse filtering, iterative and adaptive inverse filtering, and mixed-phase decomposition using synthetic signals and a large real-speech corpus.
Source: §6 · pp. 21–22Results
Closed-phase inverse filtering and complex cepstrum decomposition perform best on clean synthetic and real speech, while iterative and adaptive inverse filtering performs best in adverse noisy conditions.
Source: §4.4 · pp. 17–18, §6 · p. 22Takeaways & Limitations
NAQ, H1-H2, and HRF can discriminate voice qualities, with closed-phase inverse filtering and complex cepstrum decomposition showing similar, slightly favorable real-speech separation.
Source: §5 · p. 21, §6 · p. 22Closed-phase inverse filtering can be constrained when the closed phase is difficult to locate or too short at high pitches for reliable filter estimation.
Source: §2.1.1 · p. 5
Abstract
from arXiv · showhide
Source-tract decomposition (or glottal flow estimation) is one of the basic problems of speech processing. For this, several techniques have been proposed in the literature. However studies comparing different approaches are almost nonexistent. Besides, experiments have been systematically performed either on synthetic speech or on sustained vowels. In this study we compare three of the main representative state-of-the-art methods of glottal flow estimation: closed-phase inverse filtering, iterative and adaptive inverse filtering, and mixed-phase decomposition. These techniques are first submitted to an objective assessment test on synthetic speech signals. Their sensitivity to various factors affecting the estimation quality, as well as their robustness to noise are studied. In a second experiment, their ability to label voice quality (tensed, modal, soft) is studied on a large corpus of real connected speech. It is shown that changes of voice quality are reflected by significant modifications in glottal feature distributions. Techniques based on the mixed-phase decomposition and on a closed-phase inverse filtering process turn out to give the best results on both clean synthetic and real speech signals. On the other hand, iterative and adaptive inverse filtering is recommended in noisy environments for its high robustness.
1. Introduction
Separating glottal and vocal-tract contributions matters for modeling phonation and applications including synthesis, pathology detection, speaker recognition, and emotion analysis. This study compares three estimation methods using synthetic signals and connected real speech, addressing limitations in prior evaluation coverage.
1. Introduction
- Glottal and vocal-tract contributions have different dynamics, so separating them is a preliminary condition for studying glottal-based vocal effects.The paper notes that these effects may be segmental or controlled at a suprasegmental level.Source: §1 · p. 2
- Estimating the glottal source from speech alone is a blind-separation problem because neither the vocal-tract nor glottal contribution is observable.Natural speech also lacks a target reference signal for quantitative assessment.Source: §1 · p. 2
- Prior evaluations commonly used synthetic speech or a few sustained vowels, and comparative studies were rare.The paper cites Sturmel et al. (2007) as an example of a comparative study.Source: §1 · p. 2
- The study compares closed-phase inverse filtering, iterative and adaptive inverse filtering, and mixed-phase decomposition.Synthetic signals provide the original glottal flow for objective measures; real-speech tests examine voice-quality discrimination on a large database.Source: §1 · p. 3
- The experiments examine method sensitivity to factors affecting separation, robustness to noise, and voice-quality information in estimated glottal sources.The real-speech analysis includes tensed, modal, and soft voice qualities.Source: §1 · p. 3
2. Glottal Source Estimation
The section reviews inverse-filtering and mixed-phase approaches for estimating glottal contributions from speech. It describes their core separation mechanisms and practical conditions, including closed-phase estimation and phase-based decomposition.
2. Glottal Source Estimation
- Glottal source estimation targets the voiced excitation of the vocal tract; a glottal cycle includes opening, return, and closure phases, with the Glottal Closure Instant marking the open-to-return boundary.The paper often combines lip radiation with the glottal component and works with the glottal-flow derivative.Source: §2 · p. 3, Figure 1 · p. 4
2.1. Methods based on Inverse Filtering
- Inverse filtering estimates a vocal-tract model, then removes its contribution from speech to obtain the glottal flow.The reviewed approaches differ in how they estimate the vocal tract.Source: §2.1 · p. 4
2.1.1. Closed Phase Inverse Filtering
- Closed-phase inverse filtering estimates the vocal-tract spectral envelope during glottal closure, when subglottal-cavity effects are minimized.Its accuracy depends on locating the closed phase; at high pitches, that interval may be too short for reliable filter estimation.Source: §2.1.1 · p. 5
2.1.2. Iterative and/or Adaptive Inverse Filtering
- Iterative and adaptive approaches refine glottal and vocal-tract estimates; this study uses IAIF as the representative method.IAIF iteratively refines both components.Source: §2.1.2 · p. 6
2.2. Mixed-Phase Decomposition
- Mixed-phase decomposition separates speech into minimum-phase and maximum-phase components, with the latter attributed to the glottal contribution.The vocal-tract response and glottal return phase are treated as minimum-phase, while the glottal open phase is maximum-phase.Source: §2.2 · pp. 6–7
2.2.1. Zeros of the Z-Transform (ZZT)
- In ZZT decomposition, roots outside the unit circle represent the maximum-phase glottal open phase, while roots inside mainly represent the vocal-tract response.The unit circle serves as the discriminant boundary.Source: §2.2.1 · pp. 7–8
2.2.2. Complex Cepstrum Decomposition (CCD)
- Complex cepstrum decomposition estimates the glottal contribution by retaining negative quefrency indices, where an anticausal signal has nonzero values.The quefrency origin separates the components.Source: §2.2.2 · pp. 8–9
2.2.3. Equivalence between ZZT and CCD
- ZZT and CCD are functionally equivalent ways to separate minimum- and maximum-phase components, but CCD is 10 to 100 times faster at 16 kHz.CCD relies on FFT and IFFT operations, whereas ZZT requires factoring high-order polynomials; the study uses CCD for its higher speed.Source: §2.2.3 · p. 10
3. Glottal Source Parametrization
The section reviews time- and frequency-domain parameters for characterizing estimated glottal signals. It describes measures of waveform timing and shape, pulse amplitude, spectral characteristics, and harmonic content.
3. Glottal Source Parametrization
- Glottal-source parameters compactly characterize estimated signals using features from both time and frequency domains.Some of these features are used in the paper’s experiments.Source: §3 · p. 10
3.1. Time-domain features
- Time-domain features describe waveform shape and timing; QOQ measures the relative open time using the portion of a pitch period when flow exceeds half its peak-to-trough range.QOQ was found to reflect phonation changes better than tested Oq variants.Source: §3.1 · pp. 10–11
- NAQ characterizes the glottal closing phase as peak glottal flow divided by the minimum flow derivative, normalized by the fundamental period.Prior work cited in the paper found NAQ robust and efficient for separating phonation types.Source: §3.1 · p. 11
- Glottal estimates can also be characterized with seven LF parameters and five energy coefficients across cycle subsegments, or by fitting a model in the time domain.These approaches characterize coarse and fine structure, respectively.Source: §3.1 · p. 11
3.2. Frequency-domain features
- Frequency-domain features include glottal-formant frequency and bandwidth, spectral model-fit measures, and H1−H2, HNR, and HRF measures of voice quality or harmonic content.HRF is the sum of harmonic amplitudes divided by the amplitude at the fundamental frequency.Source: §3.2 · pp. 11–12
4. Experiments on Synthetic Speech
The synthetic-speech experiments systematically vary glottal and vocal-tract parameters and noise, then compare three source-estimation methods using glottal-parameter error rates and spectral distortion. Overall, complex cepstrum decomposition performs best on synthetic signals, while the methods differ in their sensitivity to noise, pitch, and the vocal tract.
4. Experiments on Synthetic Speech
- The protocol varies LF waveform parameters, 14 vowel filters, and SNR across slightly more than 250,000 synthetic-speech experiments, evaluating NAQ and QOQ error rates and spectral distortion.The synthetic tests focus on regular, non-pathological voices; less regular phonation is outside the study's scope.Source: §4 · pp. 12–14, Table 1 · p. 13
4.2. Sensitivity to Fundamental Frequency
- The study examines performance measures as a function of fundamental frequency, but the supplied passages do not report the outcome of that comparison.The synthetic-speech analysis considers pitch as one factor affecting decomposition quality.Source: §4 · p. 14, Figure 5 · p. 15
4.1. Robustness to Additive Noise
- Below 40 dB SNR, IAIF outperforms the other methods, while CCD and CPIF degrade particularly as noise increases.The authors attribute CCD's sensitivity to noise-related phase changes and CPIF's to errors in estimating the spectral envelope during the closed phase.Source: Figure 4 · p. 14, §4.1 · p. 15
4.3. Sensitivity to Vocal Tract
- For all methods, spectral-distortion performance degrades as F1 decreases; the effect is particularly strong for CPIF and IAIF, while CCD is only slightly affected.The authors link lower F1 to a longer vocal-tract impulse response and greater overlap between source and filter components.Source: §4.3 · p. 16, Figure 6 · p. 16
4.4. Conclusions on Synthetic Speech
- Complex cepstrum decomposition has the highest overall synthetic-speech performance; CPIF better estimates QOQ in clean, especially low-pitched speech, while IAIF performs worst in clean conditions but outperforms other methods in adverse noisy conditions.The study notes that Fg sensitivity was not explicitly measured, and less regular voices such as jittery, shimmered, or breathy voices were not addressed.Source: §4 · p. 13, §4.4 · pp. 17–18
5. Experiments on Real Speech
The real-speech experiments compare three estimation techniques on connected speech using glottal-feature distributions to assess voice-quality discrimination. CPIF and CCD perform similarly and better than IAIF overall, with CCD slightly favored.
5. Experiments on Real Speech
- The real-speech comparison extends beyond sustained vowels, using De7’s loud, modal, and soft speech and feature distributions for parametric evaluation.De7 includes about 50 minutes per voice quality; the analysis uses NAQ, H1-H2, and HRF.Source: §5 · pp. 18–19
- On the example vowel, CPIF and CCD estimates resemble expected glottal-flow shapes, whereas IAIF shows residual ripples; loud-voice GCI excitation is stronger than soft-voice excitation.Source: §5 · pp. 18–19, Figure 7 · p. 19
- All three methods show significant phonation-type histogram differences, but IAIF with H1-H2 is the notable exception to convincing separation.Source: §5 · pp. 19–20, Figure 8 · p. 20
- With stronger vocal effort, NAQ and H1-H2 decrease while HRF increases; CPIF and CCD produce strongly similar histograms.Source: §5 · p. 20, Figure 8 · p. 20
- CCD gives the strongest separation in seven of nine feature/phonation-pair comparisons, while CPIF leads in two; IAIF is generally lower.CPIF leads for loud-modal separation with NAQ and HRF.Source: §5 · p. 21, Figure 9 · p. 21
- NAQ has the highest Jensen-Shannon distance, while H1-H2 and HRF carry comparable discrimination; loud-soft distances exceed the other phonation-pair distances.Source: §5 · p. 21, Figure 9 · p. 21
6. Conclusion
The study compares glottal-flow estimation techniques using synthetic and real speech, including connected speech rather than only sustained vowels. CPIF and CCD perform similarly and rather well, while the conclusion qualifies their real-speech evidence as clean TTS recordings and calls for further testing on noisy signals.
6. Conclusion
- CPIF and CCD have similarly high effectiveness, with a slight preference for CCD; IAIF is lower in most tests but comparatively more effective at very low SNR.The synthetic-test ranking changes markedly with SNR, and CCD shows rather low robustness.Source: §6 · p. 22
- The usefulness of NAQ, H1-H2, and HRF for glottal-flow parameterization is confirmed, as is their effectiveness for distinguishing voice qualities.Source: §6 · p. 22
- Real-speech tests used clean TTS recordings, supporting CCD for expressive TTS, while noisy-signal applications require further testing.Source: §6 · p. 22
- The real-speech evaluation used a large corpus and did not restrict analysis to sustained vowels; voice-quality separability served as its measure because real glottal-flow references were unavailable.Source: §6 · p. 22