Phonation Excitation Information

• Phonation excitation information is a measure of the dynamic and spectral properties of the glottal source that energizes voiced speech.
• It integrates physical modeling, signal processing, and computational simulations to enhance applications like synthetic speech, pathology detection, and emotion recognition.
• Advanced methods combine physical equations, CFD simulations, and pitch-based synthesis to achieve robust and precise voice quality analysis.

Phonation excitation information is an essential construct in speech science, quantifying the physical, physiological, and signal-level features of the glottal source that energize voiced speech. It captures the dynamical attributes and spectral characteristics rooted in vocal-fold vibration and its interaction with subglottal and vocal-tract aerodynamics. Phonation excitation information underpins voice quality, sonority discrimination, pathology detection, synthetic speech generation, and emotion recognition. Approaches span explicit physical modeling, direct physiological measurement, and algorithmic extraction from acoustic signals.

1. Physical and Mathematical Foundations of Phonation Excitation

The classical source–filter framework models voiced speech as the convolution of a glottal excitation waveform $g(t)$ and a vocal-tract filter $h(t)$, yielding output acoustic pressure $p_{out}(t)$ and spectral product $P_{out}(\omega) = G(\omega)H(\omega)$ (Mohapatra et al., 2018). Modern research demonstrates two-way source–filter coupling: the vocal tract's impedance modulates transglottal pressure $\Delta p(t)$, actively reshaping the vocal-fold dynamics and the time-varying glottal flow $U(t)$, with pronounced impact on spectral tilt, harmonic richness, and phonation bifurcations.

Lumped-mass biomechanical models, such as the one-mass and two-mass vocal-fold systems, are governed by ODEs that balance tissue elasticity, damping, aerodynamic force, and acoustic loading (Assaneo et al., 2013): $$m_i \ddot x_i + b_i(x_i)\dot x_i + K_i(x_i) + k_c(x_i - x_j) = f_i(P_s, x_i, x_j)$$ where $f_i$ draws from Bernoulli + viscous loss relationships, the glottal area $A_g(x)$, and tissue parameters. Nonlinear stiffness and collision models yield amplitude-dependent resonance and realistic isofrequency pressure–pitch relations.

Low-order flow models describe glottal excitation in terms of time-varying flow inertance $C_{iner}(t)$ and viscous/separation losses $R_{loss}(t)$ (Malinen, 2019): $C_{iner}(t) U'(t) + R_{loss}(t) U(t) = p_s(t)$ with specific constructions for rectangular and wedge-shaped geometries. Acoustic inertance and impedance transfer functions from Webster's lossless equation provide rigorous connection to tube and vocal-tract boundary conditions.

2. Signal-Derived Measures and Computational Extraction

Direct extraction of phonation excitation from speech signals leverages glottal-synchronous cues, linear prediction, and envelope measurement. In "Sonority Measurement Using System, Source, and Suprasegmental Information," phonation excitation information is operationalized as a Hilbert-envelope peak-to-side-lobe ratio $f_6(i_k)$ of the linear-prediction residual at glottal-closure instants (Sharma et al., 2021):

• LP residual $e[n] = s[n] - \sum_{k=1}^{p} a_k s[n-k]$
• Hilbert envelope: $H_e[n] = \sqrt{e[n]^2 + e_h[n]^2}$
• Excitation feature: $f_6(i_k) = P / \mu$, where $P$ is the envelope peak at GCI, $\mu$ is the mean envelope level in the 2–3 ms post-GCI region

This unipolar amplitude curve provides robustness against noise and strong correspondence with physiological measures (e.g., DEGG peaks). In the structure of the feature vector for sonority, the excitation feature carries a weight ($w_6 \approx 0.186$) comparable to suprasegmental and system features.

Phase-based representations (group delay, chirp group delay, mixed-phase decomposition) yield complementary information about timing and glottal irregularity (Drugman et al., 2020): $\tau(\omega) = -\frac{d\theta(\omega)}{d\omega}$ Relative frame-to-frame differences in group delay and anticausal (glottal open-phase) time constants (e.g., $T_1, T_2$) are directly linked to excitation anomalies, voice quality, and disease discrimination. Chirp group delay and modified group delay offer enhanced robustness in practice.

Quantum Vocal Theory of Sound (QVTS) models phonation as an observable in a two-state Hilbert space, enabling compact encoding of excitation via the Pauli–z operator's expectation $\langle \sigma_z \rangle(t)$ and energy via a phonation Hamiltonian (Rocchesso et al., 2020). Signal processing implements "quantum measurement" using framewise Harmonic-Plus-Stochastic (HPS) extractors, mapping spectral saliency to the excitation coefficient.

3. Direct Numerical, CFD, and Aeroacoustic Modeling

High-fidelity multimodal phonation simulations operate in full 2D/3D, directly coupling multi-layered viscoelastic vocal-fold finite element models, compressible or incompressible Navier–Stokes solvers, and acoustic propagation via finite element techniques (Saurabh et al., 2020, Schoder, 2023, Lasota et al., 2023). The fundamental equations encode:

• Compressible flow: mass, momentum, and energy conservation (Navier–Stokes)
• Vocal-fold motion: finite-strain stress-strain relations in the anatomical layers
• FSI coupling: kinematic continuity and traction equilibrium at the glottal boundary

Excitation signals are represented as $\frac{\partial p}{\partial t}$ mapped from the CFD or FSI region, forming the sole source term in the linearized Perturbed Convective Wave Equation (PCWE) for acoustics: $$\frac{1}{c^2} \frac{\partial^2 \psi_a}{\partial t^2} - \nabla^2 \psi_a = -\frac{1}{\rho_0 c^2} \frac{\partial p}{\partial t}$$ (Schoder, 2023). Anisotropic Minimum Dissipation (AMD) LES models preserve fine structure and vorticity in glottal flow, boosting high-frequency content and formant amplitude over standard eddy-viscosity closures (Lasota et al., 2023). Quantitative results reveal SPL and spectral characteristics in agreement with experimental formant and vowel data.

4. Pitch Pulse-Based, Period-Synchronous Voicing Analysis and Synthesis

Pitch pulse segmentation and period-synchronous analysis enables fine-grained characterization of excitation magnitude, phase, and temporal variability within and between vowels (Ferreira et al., 7 Jun 2025). For each pitch period, harmonic magnitudes $A_\ell$ and shift-invariant phase features ("NRD") are extracted from DFT representations. Three synthesis methods translate these models into continuous or hybrid frequency/time-domain reconstructions:

Method	Synthesis Domain	Key Characteristic
FRE	Frequency	Overlap–add, "perfect-reconstruction"
TIM	Time & Frequency	Pitch-period pulse concatenation
GLO	Physiological	Per-pulse glottal LF-model + filtering

Objective and subjective evaluation demonstrates that maintaining period-synchronous control and explicit glottal modeling preserves both naturalness and spectral detail, particularly for co-articulated running speech.

5. Measurement in Physiological Modalities, Applications in SER and Pathology

Direct physiological measurement of phonation excitation employs electroglottography (EGG) to record vocal-fold contact dynamics, with signal features (e.g., open quotient, contact quotient, spectral slope) reflecting excitation strength (Zhang et al., 11 Nov 2025). Feature extraction via toolkits like openSMILE abstracts physiological signals into large-dimensional descriptors suitable for downstream classification. Estimated glottal flow via Iterative Adaptive Inverse Filtering (IAIF) can substitute for EGG in cases where physical measurement is infeasible.

Speech emotion recognition (SER) and voice pathology detection (VPD) benefit substantially from the inclusion of phonation excitation information: combining excitation-derived features with standard acoustic modalities yields consistent accuracy improvements (≈1–4%), regardless of whether excitation is directly measured or algorithmically estimated. Phase-based features, especially chirp group delay and mixed-phase time constants, demonstrate superior discrimination for VPD and complement magnitude-based routes (Drugman et al., 2020).

6. Robustness, Limitations, and Future Directions

Phonation excitation information exhibits strong resilience to moderate noise and analytical artifacts due to synchronization with physiologically motivated events (GCIs, EGG peaks, pulse segmentation), LP residual flattening, and multidomain feature extraction (Sharma et al., 2021, Zhang et al., 11 Nov 2025). However, underlying performance hinges critically on precise GCI detection, correct LP order selection, cross-modal alignment (in SER), and domain-specific inversion accuracy (IAIF for EGG estimation).

Current limitations include the restricted application in unvoiced speech regions, incomplete decompositions of glottal source impedance under nonstationary conditions, and the practical complexity of obtaining ground-truth physiological signals. Emerging opportunities reside in multimodal fusion architectures, end-to-end neural inverse models for glottal flow (beyond IAIF), and fully coupled FSI-CFD-aeroacoustic simulations for high-detail phonation analysis.

Phonation excitation information thus remains central to both foundational and applied speech science, demanding rigorous modeling, careful signal processing, and multi-scale validation for robust real-world deployment.