PaperlayerView exact revision on arXiv

Source-linked AI summary

Glottal Source Processing: from Analysis to Applications

Thomas Drugman, Paavo Alku, Abeer Alwan, Bayya Yegnanarayana

arXiv:1912.12604v1cs.SDcs.CLeess.AS

TL;DR

Glottal source analysis could complement conventional vocal-tract features, but extracting it from speech is complex and requires precise pitch or GCI information. This review surveys analysis methods and applications, and concludes that glottal information can be integrated into several voice-technology applications, despite unresolved practical challenges.

  • Problem

    Glottal features may complement vocal-tract features, but extracting them from speech is difficult, and many GCI and glottal-estimation methods depend on correct polarity.

    Source: §1 · p. 2, §3.2 · p. 9

  • Method

    The review surveys pitch tracking, GCI detection, glottal-flow estimation and parameterization, and integration into voice-technology applications.

    Source: §6 · p. 29

  • Results

    The review describes glottal-based information integrated into speech synthesis, speaker recognition, biomedical applications, and expressive speech analysis.

    Source: §6 · p. 29

  • Takeaways & Limitations

    Glottal excitation is expected to complement vocal-tract features, but analysis is more complex and requires precise F0 and/or GCI locations.

    Source: §6 · p. 29

    GCI detection in practical environments and detection of minor within-cycle excitations remain challenges.

    Source: §3.3 · p. 15

Abstract

from arXiv · show

The great majority of current voice technology applications relies on acoustic features characterizing the vocal tract response, such as the widely used MFCC of LPC parameters. Nonetheless, the airflow passing through the vocal folds, and called glottal flow, is expected to exhibit a relevant complementarity. Unfortunately, glottal analysis from speech recordings requires specific and more complex processing operations, which explains why it has been generally avoided. This review gives a general overview of techniques which have been designed for glottal source processing. Starting from fundamental analysis tools of pitch tracking, glottal closure instant detection, glottal flow estimation and modelling, this paper then highlights how these solutions can be properly integrated within various voice technology applications.

1. Introduction

Glottal source characteristics contribute to voice quality, but identifying and extracting the features behind different qualities from speech is difficult. This review introduces key glottal features and the challenges of measuring and modeling them.

1. Introduction

  • Listeners perceive subtle changes in vocal-fold vibration, but identifying which features produce a given voice quality and extracting them from speech are difficult.High-speed photography cannot capture natural sound production with a time-varying vocal tract; EGG indicates impedance changes but not the pressure changes producing speech.Source: §1 · p. 2
  • Physical or analytical glottal-vibration models help explain feature significance, but do not extract actual vibration features from speech or identify those tied to emotion or specific sounds.The review distinguishes these models from methods needed to recover speech features.Source: §1 · p. 3
  • Phonation includes modal and non-modal types; although processing usually focuses on types with vocal-fold vibration, other types also produce phonemically significant sound units.Modal voice is mostly neutral with little period-to-period variation and significant excitation around the GCI.Source: §1 · p. 3
  • Glottal source processing must characterize impulse-like excitation, closing-phase behavior, and secondary excitations, not only locate glottal closure instants (GCIs).The impulse's sharpness indicates strength or loudness; closing behavior and secondary excitations also contribute to voice quality.Source: §1 · pp. 3–4
  • Instantaneous fundamental frequency (F0) and glottal-wave shape, including opening, open, closing, and closed phases, are important features of voice quality.The review aims to provide a comprehensive account of glottal source processing.Source: §1 · p. 4
  • The review covers glottal-production physiology, pitch and GCI analysis, glottal-wave characterization, applications, and practical-environment challenges.The application scenarios include speech synthesis, speaker recognition, biomedical applications, and expressive speech analysis.Source: §1 · p. 4

2. Physiology of glottal production

Glottal production converts airflow from the lungs into an excitation that the vocal tract filters. The section relates this process to voice-source models that differ in representation, detail, and practical complexity.

2. Physiology of glottal production

  • Vocal-fold vibration modulates airflow from the lungs into an excitation signal, which the oral, nasal, and pharyngeal cavities shape before it radiates from the lips.Voiced-sound cycles arise as subglottal pressure opens the folds and tissue elasticity and aerodynamic forces help close them; cycle duration is T0 and frequency is F0.Source: §2 · p. 5
  • Source models represent glottal flow or its first derivative with differing complexity, detail, and computational efficiency.The three-parameter Rosenberg model represents opening and closing separately; LF uses four parameters and is common in synthesis, while six-parameter FL adds detail but is harder to use, and R++ is more efficient with perceptual equivalence to LF.Source: §2 · p. 6
  • EE1 permits separate adjustment of opening and closing slopes, while EE2 redefines opening and closing speeds for lower computational complexity, faster waveform generation, and more accurate pulse-shape manipulation.EE2 has also been used for automatic glottal-flow estimation from acoustic speech.Source: §2 · p. 6

3. Tools for Synchronization

Glottal-source analysis uses pitch-period-scaled windows and requires GCI position; this section reviews pitch tracking, speech-polarity detection, and GCI detection as synchronization tools.

3.1. Pitch Tracking

  • Pitch trackers generate F0 candidates from autocorrelation, cepstral, cross-correlation, modulation, or excitation cues, with multi-band methods reconciling candidates across bands.F0 is the inverse of the excitation period during voicing.Source: §3.1 · pp. 7–8
  • SRH exploits LP-residual harmonic structure for pitch and voicing decisions, and was particularly robust to additive noise, improving adverse-condition performance over six state-of-the-art techniques.SAFE instead uses learned models of noise effects on SNR-spectrum peaks to evaluate F0 candidates.Source: §3.1 · p. 8

3.2. Speech Polarity Detection

  • Polarity estimators examine glottal-waveform gradient sign, harmonic phase relations, oscillating-moment phase shifts, or skewness in excitation signals.The methods reviewed include GSGW, Phase Cut, RPS, OMPD, and RESKEW.Source: §3.2 · p. 9
  • Correct polarity matters because errors can severely impair GCI detection and glottal-source estimation; RESKEW outperformed the compared techniques on 10 corpora and was more efficient and robust.The compared robustness dimensions were noise and reverberation.Source: §3.2 · pp. 9–10

3.3. Glottal Closure Instant Detection

  • Zero-frequency filtering reduces vocal-tract resonances and locates GCIs at negative-to-positive zero crossings; it also works for aperiodic impulses and additive or babble noise.With channel degradation, the method first uses the Hilbert envelope of the LP residual.Source: §3.3 · p. 13
  • Group-delay methods locate a GCI when the average group delay is zero; DYPSA refines candidate instants using dynamic programming.The zero value corresponds to the GCI at the center of the analysis segment.Source: §3.3 · pp. 13–14
  • LP-residual methods exploit reduced predictability or peaks near GCIs, but incomplete resonance cancellation and high-frequency SNR loss can obscure the impulse-like behavior.The residual’s Hilbert envelope can produce unambiguous peaks around GCIs in clean speech.Source: §3.3 · p. 14
  • Energy and time-frequency methods require block processing, while combined approaches include YAGA and SEDREAMS; GCI research also spans phonation types, paralinguistic signals, and speech-analysis applications.SEDREAMS refines GCI locations in the LP residual, while YAGA combines inverse filtering, wavelets, group delay, and dynamic programming; GCI-based work also covers voice quality and applications such as enhancement and speaker analysis.Source: §3.3 · pp. 10, 11, 14, 15

4. Glottal Source Estimation and Parameterization

This section reviews methods for estimating glottal flow from audio and representing the estimated source with parameters. It covers inverse filtering and time- and frequency-domain parameterization.

4.1. Glottal Source Estimation

  • Closed-phase covariance analysis estimates the vocal-tract model from the glottal closed phase, but small errors in locating that phase can severely distort the excitation estimate.The method uses linear prediction with a covariance criterion and can be partly improved by refining closure-instant extraction or the modeling procedure.Source: §4.1 · p. 17
  • Joint source–tract modeling uses samples across the fundamental period, while IAIF estimates glottal excitation with a two-stage low-order LP procedure.Other approaches separate source and tract using polynomial roots or phase characteristics; the latter can estimate LF-model parameters without explicitly computing the glottal signal.Source: §4.1 · pp. 17–18

4.2. Glottal Source Parameterization

  • GIF analysis typically estimates glottal-flow waveforms first, then compresses their important properties into numerical parameters.The review groups parameterization methods into time-domain and frequency-domain techniques.Source: §4.2 · p. 18
  • Time-domain measures include ratios of opening, closing, and closed-phase durations, but inverse-filter distortions can make timing vulnerable.Threshold crossings and amplitude measures such as NAQ provide alternatives; NAQ uses flow AC amplitude and the negative peak of the flow derivative.Source: §4.2 · pp. 18–19, Figure 3 · p. 19
  • Waveform parameterization can fit artificial source models, including the LF model, or estimate physical oscillator parameters instead.A data-driven alternative applies PCA and GMMs to IAIF-estimated flow and can represent non-flat closed phases that traditional approaches fail to model.Source: §4.2 · p. 20
  • Frequency-domain measures characterize spectral tilt, harmonic structure, and the balance between harmonic and non-harmonic excitation.Examples include the alpha ratio, HRF, H1-H2, and measures of spectral decay; the harmonic/non-harmonic balance is especially relevant to disordered voices with increased aperiodicity.Source: §4.2 · pp. 20–21

5. Applicability of Glottal Source Processing

The review describes glottal-source processing in speech synthesis, speaker recognition, biomedical applications, and expressive speech analysis and synthesis. Across these areas, it reports applications of source features alongside vocal-tract, residual, or prosodic information.

5.1. Speech Synthesis

  • Using the physiological glottal source as excitation preserves glottal spectral shaping and offers voice-quality flexibility, but requires reliable estimation and modeling.A residual excitation is easier to obtain, but its flat amplitude spectrum loses glottal spectral shaping in the filter component.Source: §5.1 · p. 22
  • Glottal-source synthesis methods and residual-signal models have been reported to outperform traditional excitation in HMM-based synthesis.Residual modeling techniques yielded significantly higher naturalness than traditional pulse excitation; glottal-source approaches were also reported to outperform traditional excitation.Source: §5.1 · p. 23

5.2. Speaker Recognition

  • Glottal features alone did not outperform MFCCs in the reviewed speaker-recognition studies, but combining them with MFCCs reduced identification errors.Residual and glottal features are described as complementary to conventional MFCC features.Source: §5.2 · pp. 24–25

5.3. Biomedical Application

  • For voice-disorder detection, speech- and glottal-based features were relatively complementary, and phase features were reported as well suited to highlighting phonation irregularities.The review also reports statistically significant differentiating power for glottal-related HNR over conventional HNR in pathology discrimination.Source: §5.3 · pp. 25–26
  • In depression classification, glottal and prosodic features together discriminated better overall than prosodic and vocal-tract features.Mean jitter distinguished suicidal from nondepressed control groups, while glottal-flow-spectrum slope distinguished all three groups considered.Source: §5.3 · p. 26
  • Alaryngeal speech resynthesis uses synthetic excitation, with reported improvements from glottal-waveform replacement and F0 smoothing, and reduced perceived breathiness and harshness in a repair system.For the female alaryngeal words studied, replacing the glottal waveform and smoothing and raising F0 produced the most significant enhancement; increasing average F0 alone was less dramatic.Source: §5.3 · pp. 26–27

5.4. Expressive Speech Analysis and Synthesis

  • Glottal measures have been applied to emotion classification and expressive-style analysis, including emotion pairs with similar prosody.Single glottal features reduced classification error for 24 emotion pairs compared with pitch or energy; glottal and usual voice-quality features were also reported useful for emotional-valence detection.Source: §5.4 · p. 27
  • Glottal-flow parameters in Lombard speech changed significantly and coherently with noise type and level; greater articulation effort was associated with higher pitch and more high-frequency glottal flow.The review also describes transforming residual excitation as important for identifying emotions not fully conveyed through prosody.Source: §5.4 · p. 28

6. Conclusion

The review covers glottal-source processing from synchronization and flow estimation to applications, while noting that analysis is complex, unvalidated for glottal flow, and not yet robust in realistic speech conditions.

6. Conclusion

  • The review spans pitch tracking and GCI detection, glottal-flow estimation and parameterization, and integration into four voice-technology application areas.Source: §6 · p. 29
  • Glottal excitation is expected to complement conventional vocal-tract features, but its analysis requires precise pitch and/or GCI information.Source: §6 · p. 29
  • Glottal-flow estimation is a blind-separation problem because neither the vocal-tract response nor the glottal contribution is observable, leaving no reference for validating estimates.Source: §6 · p. 29
  • Current algorithms are sensitive and weakly robust, with further work needed for noisy or reverberant connected speech, high-pitched voices, expressive speech, and conversational data.Source: §6 · p. 29
Loading 1912.12604v1…