Why One Number Isn't Enough: Dimension-Specific Voice Quality Indices

Perceptual voice assessment is explicitly multidimensional. The CAPE-V asks raters to score six qualities on separate 100 mm visual analog scales; GRBAS separates grade, roughness, breathiness, asthenia, and strain. The clinical rationale is straightforward: different qualities point toward different mechanisms, and different mechanisms call for different interventions. Breathiness suggests incomplete glottal closure. Roughness suggests irregular vocal fold vibration. Strain suggests hyperfunction.

On the acoustic side, the landscape has been lopsided. AVQI (Maryn et al., 2010) combines six parameters into a single severity score and has accumulated validation across more than a dozen languages. ABI (Barsties von Latoszek et al., 2017) targets breathiness with nine parameters. CSID (Awan et al., 2016) estimates overall severity from spectral and cepstral measures. All three are valuable. None of them tells you how rough or how strained a voice is.

This is not because roughness and strain lack acoustic correlates—individual parameters like jitter, shimmer, and spectral tilt have long been associated with them. It is because nobody had published a data-driven composite index for either dimension. Single parameters are noisy and unreliable in isolation; the whole point of the multiparametric approach is that combinations outperform any single measure. Roughness and strain simply never got their combination.

The four indices are linear combinations of acoustic parameters extracted from a sustained /a/ vowel of at least three seconds. Each index outputs a score on the CAPE-V 100-point scale—the same scale clinicians already use perceptually—where higher means more deviant. Each index uses a different, deliberately small set of parameters:

Notice that GNE appears in all four indices—consistent with earlier findings that glottal-to-noise excitation carries information no other common parameter does (see our GNE guide). Period standard deviation appears in three. This sharing is not sloppiness; it reflects the real correlational structure of dysphonic voices. Perceptual dimensions overlap, and so do their acoustic footprints. The indices differ in weighting and in the parameters that are unique to each.

The methodology matters, because it is what separates these indices from an arbitrary weighting of someone's favorite parameters. The derivation used the Perceptual Voice Qualities Database (PVQD; Walden, 2022)—296 speakers with sustained vowels and CAPE-V ratings from multiple expert listeners—reduced to 284 after removing recordings with incomplete data.

Fourteen candidate parameters were extracted from each vowel: measures of perturbation (jitter, shimmer in % and dB, period standard deviation), noise (HNR, GNE, CPPS, high-frequency noise, HNR-D), spectral shape (slope, tilt, alpha ratio, H1-H2), and pitch (F0). Four right-skewed parameters were log-transformed. Then selection proceeded in two stages:

The result is that each index's parameter set has a defensible answer to the question "why these and not others?"—the answer being that the data, not the author, made the selection. The full regression coefficients, analysis code, and derivation details are in the open-access paper and its public repository (Lucero, 2026).

Classification accuracy was assessed against the standard criterion of CAPE-V ≥ 10 on each dimension—the threshold for clinically relevant deviation—using area under the ROC curve (AUC) with repeated cross-validation:

For context, AVQI and CSID rescored on the same 284 speakers achieved AUCs of 0.827 and 0.787 against overall severity. The new Severity Index's 0.870 outperformed both on this sample—but that comparison deserves an honest asterisk, which brings us to the next section.

Every measurement tool earns trust by being clear about its boundaries. Four boundaries matter here:

External validation on human voices takes time—new samples, new raters, new studies. But there is a complementary check available now: physics-based simulation. Using a computational model of vocal fold vibration, we can synthesize voices in which a single mechanism is perturbed at a time—aspiration noise alone, jitter alone, vibratory asymmetry alone, hyperfunctional strain alone—and ask whether each dimension index responds selectively to its matching mechanism.

The logic is simple: if the Breathiness Index is really tracking breathiness, it should respond disproportionately when aspiration noise is dialed up and other mechanisms are held fixed—something impossible to arrange with human voices, where mechanisms co-occur. Preliminary results from this mechanism-selectivity analysis, using our physics-based simulator across controlled perturbation series in male and female voice configurations, show the predicted selectivity pattern in the large majority of test conditions, with one intriguing sex-specific exception under asymmetric vibration that we are investigating further. A full report is in preparation.

The PhonaLab implementation is deliberately simple. Upload or record a sustained /a/ of at least three seconds at a sampling rate of 22,050 Hz or higher (44,100 Hz recommended). The tool returns the four scores, each with:

• A threshold interpretation against the same CAPE-V ≥ 10 criterion used in the derivation—below the threshold, or at/above it. No invented severity brackets: the paper defines one threshold, so the tool reports one threshold.
• The contributing parameters—the raw acoustic values that entered each score, so you can see why an index landed where it did. A high Breathiness score with a collapsed GNE tells a different story than one driven by low CPPS.
• Full methods and citation—every number traces to the source paper, and the analysis pipeline (Praat 6.1.38 via Parselmouth) is documented in the report footer.

One practical note: because the tool reproduces the paper's parameter extraction exactly, the raw parameter values shown in the expanders may differ slightly from those in PhonaLab's AVQI/ABI tool, which follows the Maryn & Latoszek protocol. Each tool is faithful to its own source publication—that is a feature, not a discrepancy.