Sustained Vowel Phonation (SVP) Overview

• Sustained vowel phonation (SVP) is a controlled vocal task where a single vowel is held under steady articulatory conditions to isolate phonatory function.
• Research employs varied recording protocols and feature extraction methods, enabling assessments in neurodegenerative disorders, singing, and biomechanical modeling.
• SVP supports integrated frameworks—from statistical models to deep learning and inverse biomechanical inference—for clinical diagnostics and voice synthesis validation.

Sustained vowel phonation (SVP) is the production of a single vowel in isolation under approximately stationary laryngeal and articulatory conditions, most often with instructions to sustain the sound at comfortable pitch and loudness for as long as possible. In the cited literature, SVP serves as a controlled probe of phonatory function because it reduces the linguistic planning, prosodic variation, and co-articulatory transitions that complicate running-speech analysis; for that reason it appears in clinical voice assessment, neurodegenerative-disease biomarker studies, singing-phonation classification, auditory-feedback experiments, and biomechanical simulations of voice production (Riad et al., 2020, Vashkevich et al., 2020, Ferreira et al., 7 Jun 2025).

1. Task definition and acquisition regimes

SVP is operationalized differently across subfields, but the core task remains stable: a speaker produces a held vowel on one breath while non-steady onset and offset regions are either excluded analytically or physically trimmed. In Huntington’s disease (HD), subjects inhaled deeply and produced a single /a/ at a comfortable, constant pitch and loudness until voluntary termination or breath exhaustion; recordings used a ZOOM H4n Pro microphone at 44.1 kHz and 16-bit resolution under quiet clinical conditions, with typical durations of approximately 17 s for controls, 16 s for premanifest HD, and 9 s for manifest HD (Riad et al., 2020). In bulbar ALS studies, the task was similarly implemented with /a/ or /i/, recorded on smartphones with a headset microphone at 44.1 kHz, 16-bit PCM, and trimmed to remove non-steady segments; typical samples were approximately 3.7–4.1 s or about 4.1 s depending on cohort and protocol (Vashkevich et al., 2020, Vashkevich et al., 2020).

Other studies use shorter, task-specific windows. For pathological subharmonic detection, sustained vowels were represented as roughly 1.1 s recordings, with the first 0.1 s discarded to eliminate onset transients and the remainder resampled to 8 kHz (Ikuma et al., 15 Jan 2025). In singing, the Proutskova–Rhodes soprano corpus contains 763 sustained-vowel recordings spanning nine vowels and four phonation labels, originally at 44.1 kHz and downsampled to 16 kHz, with durations of roughly 2–4 s (Justus et al., 14 Feb 2026). In Parkinson’s disease (PD), SVP was treated as a “purely vocal” task with utterances typically up to 30 s, recorded at 16 kHz using a headset microphone in the Quebec Parkinson Network corpus (Plantinga et al., 14 Jul 2025).

Context	Vowels / duration	Acquisition
HD	/a/; until breath exhaustion; controls ≈ 17 s, preHD ≈ 16 s, HD ≈ 9 s	ZOOM H4n Pro, 44.1 kHz, 16-bit, quiet clinic
ALS	/a/, /i/; typical ≈ 3.7–4.1 s or ≈ 4.1 s	Smartphone + headset, 44.1 kHz, 16-bit PCM
Subharmonic detection	held vowel, ~1.1 s with first 0.1 s discarded	resampled to 8 kHz, zero-mean, unit-variance
Singing phonation modes	nine vowels, ~2–4 s	44.1 kHz downsampled to 16 kHz
PD SVP	single sustained vowel, chunks of ≤ 30 s during training	headset mic, 16 kHz, 16-bit WAV

These acquisition differences matter methodologically. Very short windows favor framewise detectors of periodic structure or subharmonicity, whereas longer recordings support perturbation measures, tremor estimation, and longitudinal monitoring. A plausible implication is that “SVP” is better understood as a family of closely related controlled-vowel protocols than as a single fixed recording standard.

2. Acoustic representations and phonatory descriptors

A major line of SVP research relies on hand-crafted descriptors of airflow insufficiency, aperiodicity, irregular vocal-fold vibration, perturbation, noise, tremor, and spectral instability. In HD, the extracted phonatory feature set included Maximum Phonation Time, First Occurrence of Voice Break, Number of Voice Breaks, Degree of Pitch Breaks, Degree of Vocal Arrests, standard deviation of fundamental frequency $\sigma F_0$, Recurrence Period Density Entropy, local jitter, local shimmer, Harmonics-to-Noise Ratio, Detrended Fluctuation Analysis, tremor indices, and means of the standard deviations of MFCCs and delta-MFCCs (Riad et al., 2020). In ALS, parallel feature families included jitter, shimmer, Directional Perturbation Factor, HNR, Glottal-to-Noise Excitation ratio, Pitch Period Entropy, Pathological Vibrato Index, formants, MFCCs, and pitch-synchronous harmonic-structure measures such as $H_p^\mu$, $H_p^\sigma$, and $\mathrm{RelH}_p$ (Vashkevich et al., 2020).

Two perturbation measures recur throughout the SVP literature:

$$J_{loc} = \frac{\tfrac{1}{N-1}\sum_{i=1}^{N-1}|T_i - T_{i+1}|} {\tfrac{1}{N}\sum_{i=1}^{N}T_i} \times 100\%$$

and

$$S_{loc} = \frac{\tfrac{1}{N-1}\sum_{i=1}^{N-1}|A_i - A_{i+1}|} {\tfrac{1}{N}\sum_{i=1}^{N}A_i} \times 100\%,$$

where $T_i$ denotes period duration and $A_i$ peak amplitude (Vashkevich et al., 2020). Such measures require reliable cycle segmentation, which several studies treat as a central technical problem rather than a preprocessing triviality.

Beyond scalar perturbation features, SVP supports richer time–frequency representations. In HD, Modulation Power Spectrum (MPS) features were computed from a high-resolution spectrogram using 100 ms windows, 1 ms hop, 50 Hz frequency bins, and log-amplitude scaling, followed by a two-dimensional Fourier transform,

$\mathrm{MPS}(u,v)=\mathcal F_{t,f}\{S(t,f)\}(u,v),$

with temporal modulations retained from $-200$ to $H_p^\mu$0 Hz and spectral modulations from $H_p^\mu$1 to $H_p^\mu$2 cycles/kHz (Riad et al., 2020). The paper interprets low temporal and low spectral modulations near the origin as reflecting stable sustained voicing and roughness, with peaks corresponding to voice breaks or tremor, while mid spectral modulations around 4 cycles/kHz at low temporal modulation are linked to harmonic spacing and roughness.

A different representation is phase-synchronous harmonic analysis. For natural vowels and whispered-speech voicing restoration, pitch pulses were segmented by aligning the zero-crossing or phase onset of the fundamental, then harmonic magnitudes $H_p^\mu$3 and phases $H_p^\mu$4 were extracted per pulse; the normalized relative delay feature was defined as

$H_p^\mu$5

The resulting sequence $H_p^\mu$6 characterizes the waveform shape of each pulse and makes explicit the distinction between the near-stationary pulse-to-pulse structure of sustained vowels and the evolving structure of co-articulated vowels (Ferreira et al., 7 Jun 2025).

3. Analytical, machine-learning, and inverse-model frameworks

SVP has been analyzed with regularized linear models, discriminant classifiers, hierarchical pipelines, fully convolutional networks, self-supervised encoders, and dynamical-system inversion. In HD, disease-stage classification used multinomial logistic regression with an ElasticNet penalty, trained on phonatory features, MPS features, or both, and evaluated by 100 repeats of random 80/20 train/test splits using accuracy and macro-averaged $H_p^\mu$7 (Riad et al., 2020). The same study used linear regression with ElasticNet to predict clinical scores from phonatory features. ALS studies used Linear Discriminant Analysis after feature selection or compact perturbation/vibrato subsets; one formulation used the decision function $H_p^\mu$8, with Fisher-optimal weights maximizing between-class separation over within-class variance (Vashkevich et al., 2020, Vashkevich et al., 2020).

At the deep-learning end, pathological subharmonic voicing was approached with fully convolutional neural networks operating directly on raw 1D audio, without explicit spectrogram or time-domain feature extraction. The FCN outputs per-frame probabilities $H_p^\mu$9 and a framewise class decision

$H_p^\sigma$0

where $H_p^\sigma$1 denotes normal periodic phonation and $H_p^\sigma$2 denotes a subharmonic order (Ikuma et al., 15 Jan 2025). Two receptive-field variants, FCN-401 and FCN-785, were studied.

A separate line of work uses pretrained speech representations as fixed feature extractors for SVP. In singing, voice2mode pooled layerwise embeddings from HuBERT and wav2vec 2.0 and classified the pooled vectors with SVM or XGBoost, showing that early layers were more effective than higher ASR-specialized layers (Justus et al., 14 Feb 2026). In PD detection, frozen ASR-based, self-supervised, and AudioSet-pretrained encoders were followed by a small binary classifier with attention pooling; the comparison included SB VocalFeats, mel-filterbank, OpenSMILE eGeMAPSv02, Whisper variants, XEUS, WavLM, HuBERT, and wav2vec 2.0 (Plantinga et al., 14 Jul 2025).

SVP also supports inverse biomechanical inference. In COVID-19 detection, each 50 ms SVP segment was mapped to parameters of a one-mass asymmetric body–cover model via the ADLES algorithm, which minimizes the residual energy between model-predicted and inverse-filtered glottal flow under ODE constraints (Ismail et al., 2020). This places SVP in a distinct methodological category: the task is not only a source of observable acoustic features but also a stable enough signal for parameter estimation in explicit dynamical systems.

4. Clinical biomarker applications

The strongest evidence base for SVP concerns voice and neurological disorders, but the cited studies do not support a uniform conclusion across diseases. In HD, manifest disease was robustly discriminated from controls by phonatory features related to airflow insufficiency and articulatory MFCC measures, with Cohen’s $H_p^\sigma$3–$H_p^\sigma$4 for control versus manifest HD comparisons. However, no single phonatory scalar feature reached significance after correction for premanifest HD versus control. Classification results were modest: phonatory features yielded $H_p^\sigma$5 and $H_p^\sigma$6, MPS yielded $H_p^\sigma$7 and $H_p^\sigma$8, and combined features yielded $H_p^\sigma$9 and $\mathrm{RelH}_p$0. For clinical-severity regression, phonatory features predicted cUHDRS with $\mathrm{RelH}_p$1 and $\mathrm{RelH}_p$2, TFC with $\mathrm{RelH}_p$3 and $\mathrm{RelH}_p$4, and TMS with $\mathrm{RelH}_p$5 and $\mathrm{RelH}_p$6; Maximum Phonation Time and First Occurrence of Voice Break were the strongest predictors (Riad et al., 2020).

ALS studies reported substantially higher classification performance, although with different feature sets and cohorts. One study based on perturbation and pathological vibrato obtained best LDA performance using $\mathrm{RelH}_p$7, reaching $\mathrm{RelH}_p$8, $\mathrm{RelH}_p$9, $$J_{loc} = \frac{\tfrac{1}{N-1}\sum_{i=1}^{N-1}|T_i - T_{i+1}|} {\tfrac{1}{N}\sum_{i=1}^{N}T_i} \times 100\%$$0, and $$J_{loc} = \frac{\tfrac{1}{N-1}\sum_{i=1}^{N-1}|T_i - T_{i+1}|} {\tfrac{1}{N}\sum_{i=1}^{N}T_i} \times 100\%$$1 (Vashkevich et al., 2020). A broader ALS feature study using /a/ and /i/, 131 features per speaker, and LASSO-selected LDA reported $$J_{loc} = \frac{\tfrac{1}{N-1}\sum_{i=1}^{N-1}|T_i - T_{i+1}|} {\tfrac{1}{N}\sum_{i=1}^{N}T_i} \times 100\%$$2, $$J_{loc} = \frac{\tfrac{1}{N-1}\sum_{i=1}^{N-1}|T_i - T_{i+1}|} {\tfrac{1}{N}\sum_{i=1}^{N}T_i} \times 100\%$$3, and $$J_{loc} = \frac{\tfrac{1}{N-1}\sum_{i=1}^{N-1}|T_i - T_{i+1}|} {\tfrac{1}{N}\sum_{i=1}^{N}T_i} \times 100\%$$4 for a 32-feature model, while a compact 5-feature model reached $$J_{loc} = \frac{\tfrac{1}{N-1}\sum_{i=1}^{N-1}|T_i - T_{i+1}|} {\tfrac{1}{N}\sum_{i=1}^{N}T_i} \times 100\%$$5 accuracy with $$J_{loc} = \frac{\tfrac{1}{N-1}\sum_{i=1}^{N-1}|T_i - T_{i+1}|} {\tfrac{1}{N}\sum_{i=1}^{N}T_i} \times 100\%$$6 sensitivity and $$J_{loc} = \frac{\tfrac{1}{N-1}\sum_{i=1}^{N-1}|T_i - T_{i+1}|} {\tfrac{1}{N}\sum_{i=1}^{N}T_i} \times 100\%$$7 specificity (Vashkevich et al., 2020).

For benign laryngeal disorders, a hierarchical framework on 15,132 recordings from 1,261 speakers used short sustained vowels /a/, /i/, and /u/ across neutral, high, low, and gliding pitch conditions. Stage 1 achieved $$J_{loc} = \frac{\tfrac{1}{N-1}\sum_{i=1}^{N-1}|T_i - T_{i+1}|} {\tfrac{1}{N}\sum_{i=1}^{N}T_i} \times 100\%$$8 test accuracy for pathological versus healthy screening, with sensitivity $$J_{loc} = \frac{\tfrac{1}{N-1}\sum_{i=1}^{N-1}|T_i - T_{i+1}|} {\tfrac{1}{N}\sum_{i=1}^{N}T_i} \times 100\%$$9, specificity $$S_{loc} = \frac{\tfrac{1}{N-1}\sum_{i=1}^{N-1}|A_i - A_{i+1}|} {\tfrac{1}{N}\sum_{i=1}^{N}A_i} \times 100\%,$$0, ROC-AUC $$S_{loc} = \frac{\tfrac{1}{N-1}\sum_{i=1}^{N-1}|A_i - A_{i+1}|} {\tfrac{1}{N}\sum_{i=1}^{N}A_i} \times 100\%,$$1, and PR-AUC $$S_{loc} = \frac{\tfrac{1}{N-1}\sum_{i=1}^{N-1}|A_i - A_{i+1}|} {\tfrac{1}{N}\sum_{i=1}^{N}A_i} \times 100\%,$$2. Stage 2 achieved $$S_{loc} = \frac{\tfrac{1}{N-1}\sum_{i=1}^{N-1}|A_i - A_{i+1}|} {\tfrac{1}{N}\sum_{i=1}^{N}A_i} \times 100\%,$$3 accuracy for etiological triage with macro-averaged ROC-AUC $$S_{loc} = \frac{\tfrac{1}{N-1}\sum_{i=1}^{N-1}|A_i - A_{i+1}|} {\tfrac{1}{N}\sum_{i=1}^{N}A_i} \times 100\%,$$4. Stage 3 achieved $$S_{loc} = \frac{\tfrac{1}{N-1}\sum_{i=1}^{N-1}|A_i - A_{i+1}|} {\tfrac{1}{N}\sum_{i=1}^{N}A_i} \times 100\%,$$5 accuracy for 9-way subtype classification with macro-averaged ROC-AUC $$S_{loc} = \frac{\tfrac{1}{N-1}\sum_{i=1}^{N-1}|A_i - A_{i+1}|} {\tfrac{1}{N}\sum_{i=1}^{N}A_i} \times 100\%,$$6 and PR-AUC $$S_{loc} = \frac{\tfrac{1}{N-1}\sum_{i=1}^{N-1}|A_i - A_{i+1}|} {\tfrac{1}{N}\sum_{i=1}^{N}A_i} \times 100\%,$$7; the hierarchical system improved over a flat 9-way baseline by $$S_{loc} = \frac{\tfrac{1}{N-1}\sum_{i=1}^{N-1}|A_i - A_{i+1}|} {\tfrac{1}{N}\sum_{i=1}^{N}A_i} \times 100\%,$$8 ROC-AUC and $$S_{loc} = \frac{\tfrac{1}{N-1}\sum_{i=1}^{N-1}|A_i - A_{i+1}|} {\tfrac{1}{N}\sum_{i=1}^{N}A_i} \times 100\%,$$9 PR-AUC (Annabestani et al., 31 Dec 2025).

In COVID-19 screening, SVP segments analyzed through vocal-fold oscillation parameters yielded ROC-AUC $T_i$0 with logistic regression when vowels were pooled, with vowel-specific logistic-regression AUCs of $T_i$1 for /a/, $T_i$2 for /i/, and $T_i$3 for /u/ (Ismail et al., 2020). In early PD detection, by contrast, SVP was markedly weaker than spontaneous-speech tasks: SB VocalFeats reached $T_i$4 F1, XEUS $T_i$5, and HuBERT Base–AS $T_i$6, whereas spontaneous picture-description speech reached approximately $T_i$7 F1 for the best model (Plantinga et al., 14 Jul 2025).

Taken together, these findings indicate that SVP is highly informative for manifest phonatory pathology and some diagnostic settings, but not universally sufficient as a standalone biomarker. That conclusion is explicit for premanifest HD and implicitly reinforced by the PD comparison between purely vocal and language-bearing tasks.

5. Singing, subharmonics, and auditory–motor control

Outside pathology triage, SVP functions as a controlled substrate for studying phonation modes and fine-grained glottal phenomena. In singing, the task has been formulated as four-way classification among breathy, neutral (modal), flow, and pressed phonation. On 763 sustained-vowel recordings from a professional Russian soprano, HuBERT embeddings from layer 5 combined with SVM achieved $T_i$8 accuracy, compared with $T_i$9 for a spectrogram baseline, $A_i$0 for mel-spectrogram, and $A_i$1 for MFCC. The study further reported a significantly negative slope in the linear fit $A_i$2, with $A_i$3 and $A_i$4, showing that classification accuracy decreased with layer depth (Justus et al., 14 Feb 2026).

SVP is also a canonical environment for subharmonic analysis because nearly periodic held vowels make departures from period-$A_i$5 oscillation measurable. In the FCN study, subharmonic orders $A_i$6, $A_i$7, and $A_i$8 were emphasized as common pathological cases. Synthetic evaluation reached $A_i$9 overall accuracy for FCN-401 and $\mathrm{MPS}(u,v)=\mathcal F_{t,f}\{S(t,f)\}(u,v),$0 for FCN-785. On real sustained-vowel recordings from the KayPENTAX Disordered Voice Database, the networks reliably detected sustained subharmonic locking even for cases not explicitly represented in training, but they tended to suppress unlocked or weak subharmonics by outputting $\mathrm{MPS}(u,v)=\mathcal F_{t,f}\{S(t,f)\}(u,v),$1; no global precision/recall numbers were reported for real data (Ikuma et al., 15 Jan 2025).

SVP also enables controlled experiments on auditory feedback and involuntary pitch regulation. Using mixtures of orthogonal sequences made from extended time-stretched pulses to frequency-modulate an auditory test signal while subjects sustained /a/, the reported compensatory response showed approximately $\mathrm{MPS}(u,v)=\mathcal F_{t,f}\{S(t,f)\}(u,v),$2 ms latency to half-peak and approximately $\mathrm{MPS}(u,v)=\mathcal F_{t,f}\{S(t,f)\}(u,v),$3 ms latency to peak, with a peak compensatory shift of approximately $\mathrm{MPS}(u,v)=\mathcal F_{t,f}\{S(t,f)\}(u,v),$4 cents. Response SNR was typically $\mathrm{MPS}(u,v)=\mathcal F_{t,f}\{S(t,f)\}(u,v),$5–$\mathrm{MPS}(u,v)=\mathcal F_{t,f}\{S(t,f)\}(u,v),$6 dB, compared with approximately $\mathrm{MPS}(u,v)=\mathcal F_{t,f}\{S(t,f)\}(u,v),$7 dB for conventional step-perturbation methods (Kawahara et al., 2021). This extends SVP from static voice assessment to dynamic characterization of the auditory–motor loop.

6. Biomechanical modeling, synthesis, and aeroacoustics

SVP is also a favored benchmark for source modeling because the task approximates quasi-steady phonation while preserving the essential nonlinearities of vocal-fold vibration. In a single-degree-of-freedom vocal-fold model fit to high-speed videoendoscopy of four normophonic subjects producing sustained /i/, the governing dynamics were written as

$\mathrm{MPS}(u,v)=\mathcal F_{t,f}\{S(t,f)\}(u,v),$8

where $\mathrm{MPS}(u,v)=\mathcal F_{t,f}\{S(t,f)\}(u,v),$9 is a Bernoulli-based aerodynamic force, $-200$0 an added flow-separation resistance during closing, and $-200$1 a structural contact force during closure (Ali et al., 27 Jun 2026). Glottal area waveforms extracted from 4,000 fps HSV via U-Net segmentation were used for particle-swarm optimization of subject-specific parameters, yielding normalized errors below $-200$2 and reported errors of $-200$3, $-200$4, $-200$5, and $-200$6 across four subjects.

At a larger scale, sustained-vowel phonation has been simulated with hybrid LES/aeroacoustic pipelines. In the AMD study, incompressible laryngeal flow was computed in OpenFOAM on a 2.2 million-cell mesh, and acoustic propagation was solved in openCFS on an acoustic mesh of approximately 33,000 hexahedral elements. Five cardinal vowels /u, i, ɑ, o, æ/ were generated with prescribed sinusoidal glottal motion at $-200$7 Hz, amplitudes $-200$8 mm, and constant subglottal pressure of $-200$9 Pa (Lasota et al., 2023). The anisotropic minimum dissipation model yielded stronger sound pressure levels at higher harmonics and especially at the first two formants than WALE, and AMD formant estimates lay within measured ranges.

SVP is equally important for analysis–synthesis work. In whispered-to-natural voice restoration, phase-based pitch-pulse segmentation was used to compare sustained and co-articulated vowels and to drive three lightweight voicing-restoration methods: frequency-domain reconstruction (FRE), combined time/frequency-domain reconstruction (TIM), and physiologically inspired glottal-pulse filtering (GLO). Objective comparisons showed that all three preserved $H_p^\mu$00–$H_p^\mu$01 formant tracks; in subjective tests using the ITU-R BS.1116 protocol and a 100-point Continuous Quality Scale, FRE scores for sustained vowels were statistically indistinguishable from the original, whereas in co-articulated vowels GLO consistently outscored TIM with $H_p^\mu$02 (Ferreira et al., 7 Jun 2025).

A plausible implication is that SVP’s relative stationarity makes it a privileged regime for validating source models, pulse segmentation schemes, and reconstruction algorithms before those methods are extended to the stronger nonstationarities of continuous speech.

7. Limitations, misconceptions, and current research directions

A recurrent misconception is that SVP is intrinsically a complete biomarker. The cited studies do not support that position. In HD, SVP alone was explicitly described as insufficient as a standalone biomarker for subclinical or premanifest disease, although it remained useful for clinical-severity prediction and rapid assessment of phonatory motor control (Riad et al., 2020). In PD, SVP behaved as a “purely vocal” task and benefited from AudioSet pretraining, but spontaneous speech substantially outperformed it, suggesting that language-bearing tasks capture additional disease-relevant information (Plantinga et al., 14 Jul 2025).

Another limitation concerns ecological validity. Synthetic subharmonic training omitted important real-world effects such as $H_p^\mu$03 variability, biphonation-style subharmonics, intermittency, and more realistic recording artifacts; the FCNs were described as strict detectors that rarely invent subharmonics but may miss short or weak ones (Ikuma et al., 15 Jan 2025). The COVID-19 study used a small clinically curated dataset of 19 subjects, noted uncontrolled recording conditions, and emphasized that exclusiveness to COVID-19 versus other respiratory conditions had not been established (Ismail et al., 2020). The bulbar ALS perturbation-and-vibrato study likewise noted the small ALS cohort of $H_p^\mu$04 and the need for validation on larger, more diverse populations and other dysarthrias (Vashkevich et al., 2020).

Several papers converge on practical recommendations. For clinical research, SVP should be recorded with high-quality microphones, controlled microphone placement, and sufficiently high sampling rate; one HD recommendation explicitly specifies $H_p^\mu$05 kHz and standardized microphone–speaker distance, while the benign-laryngeal-disorder framework uses a head-mounted condenser microphone approximately $H_p^\mu$06 cm from the mouth at $H_p^\mu$07 off-axis in a quiet clinical room (Riad et al., 2020, Annabestani et al., 31 Dec 2025). Feature extraction should combine interpretable perturbation and noise measures with richer spectral or learned representations when possible; hierarchical or regularized models were favored when interpretability and stable feature selection mattered (Annabestani et al., 31 Dec 2025, Riad et al., 2020).

Future directions stated in the cited work include longitudinal tracking, multimodal combination with connected speech or diadochokinetic tasks, improved tremor-tracking algorithms, larger repeated-measures cohorts, fine-tuning self-supervised models on multi-singer or disease-specific corpora, richer synthetic training corpora for subharmonic detection, and extensions of reduced-order biomechanical models to cycle-to-cycle variability, nonlinear tissue rheology, or source–tract coupling (Riad et al., 2020, Justus et al., 14 Feb 2026, Ikuma et al., 15 Jan 2025, Ali et al., 27 Jun 2026). This suggests that SVP is best viewed not as a replacement for broader speech assessment, but as a precisely controlled experimental and clinical primitive on which more comprehensive voice-science pipelines can be built.