CAPE-V and GRBAS for the Acoustic Era: Reconciling Perceptual and Acoustic Voice Assessment

Before mapping perceptual dimensions to acoustic parameters, it is worth being precise about what the mapping is and is not.

A CAPE-V Breathiness rating answers: How breathy does this voice sound to a trained listener? An acoustic measure of breathiness (e.g., CPPS, GNE) answers: What spectral or noise property of the signal correlates with that perception? The two are related but not identical. A voice can produce signal patterns that statistically associate with breathiness without sounding breathy in a particular sample, and a clinician can perceive breathiness from cues that do not map cleanly to any single parameter.

This distinction matters because it sets the right expectation. Acoustic analysis cannot replace the perceptual rating; it can quantify a related signal property that, in aggregate, tracks the perception well. The clinical use is reproducibility and documentation, not substitution.

When the literature reports that an acoustic parameter "predicts" CAPE-V Breathiness with R² = 0.6, this means that variation in the parameter explains roughly 60% of the variation in breathiness ratings across speakers. The remaining 40% reflects genuine perceptual information not captured by the parameter, plus rater noise.

Two implications follow. First, no single parameter—and no combination of parameters from sustained vowels—captures everything trained listeners hear. Second, the ceiling on prediction accuracy is set partly by inter-rater reliability itself: if two raters agree on Roughness only 70% of the time, no acoustic model can predict their averaged ratings perfectly.

With those caveats in place, here is what the evidence shows for each dimension.

Severity and Grade ask essentially the same question: how deviant is this voice overall? They are the most reliable perceptual dimensions and the most studied acoustically.

Top acoustic correlates (in approximate order of incremental contribution):

• Shimmer (in dB)—cycle-to-cycle amplitude irregularity
• GNE (Glottal-to-Noise Excitation Ratio)—periodic versus aperiodic glottal source energy
• HNo-6000—relative high-frequency noise energy
• CPPS—overall harmonic strength
• F0—fundamental frequency, contributing modestly

Multiparametric indices that combine several of these parameters (AVQI, CSID) achieve good agreement with overall severity ratings, with reported AUC values around 0.80–0.87 for distinguishing dysphonic from non-dysphonic voices in research databases.

Breathiness arises when the vocal folds fail to close completely during phonation, allowing turbulent airflow to escape through the glottal gap. The acoustic consequence is added aperiodic energy—noise—mixed with the periodic harmonic structure.

Top acoustic correlates:

• CPPS—decreases as harmonic structure is degraded by aspiration noise
• GNE—directly quantifies the periodic-to-aperiodic ratio at the glottal source
• HNo-6000—captures high-frequency noise energy where aspiration is most prominent

Breathiness is the dimension where acoustic prediction works best. Three carefully chosen parameters can predict breathiness ratings nearly as well as nine-parameter indices, suggesting that the acoustic signature of incomplete glottal closure is relatively concentrated in a few measures.

Older clinical training sometimes emphasized HNR as the breathiness measure. HNR does correlate with breathiness, but it requires reliable F0 tracking and degrades when the signal becomes severely aperiodic. CPPS does not require period-by-period tracking and is more robust in moderate-to-severe cases. For most clinical purposes, CPPS and GNE outperform HNR for breathiness.

Roughness is associated with irregular vocal fold vibration: asymmetric oscillation, aperiodic cycle-to-cycle variation, subharmonic patterns, or diplophonia. It is perceptually distinct from breathiness—a rough voice is not necessarily breathy, and vice versa—but the two often co-occur in clinical samples.

Top acoustic correlates:

• Shimmer (in dB)—directly reflects the cycle-to-cycle amplitude instability of irregular vibration
• H1-H2 (first vs. second harmonic amplitude)—reflects the spectral shape of the glottal source; flatter spectra (smaller H1-H2) accompany irregular vibration
• GNE—aperiodic excitation rises with vibratory irregularity
• Jitter (in log form)—captures cycle-to-cycle frequency irregularity

Prediction accuracy for roughness is meaningfully lower than for breathiness. This appears to be substantive, not a measurement artifact: roughness encompasses a heterogeneous set of vibratory abnormalities, from mild irregularity to subharmonic breaks and diplophonia, that linear acoustic measures from sustained vowels capture only partially.

Strain reflects effortful, hyperfunctional phonation: increased laryngeal muscle tension, firm glottal closure, and elevated subglottal pressure. It is the most clinically nuanced of the four dimensions and the hardest to predict acoustically.

Top acoustic correlates:

• HNR—often elevated in strained voice (clean periodicity can coexist with hyperfunction)
• F0—tends to rise with increased longitudinal tension of the vocal folds
• Spectral tilt—less negative tilt (more high-frequency energy) accompanies pressed phonation
• H1-H2—decreases with firmer glottal adduction

Strain prediction works less well than the other dimensions for two reasons. First, strain is perceptually complex—clinicians integrate cues from voice onset, register transitions, and effort that go beyond steady-state vowel acoustics. Second, hyperfunction can produce cleaner signals on some measures (higher HNR, lower jitter), which is the opposite of what most acoustic indices were designed to detect. A perfectly periodic but pressed voice will not look "abnormal" on traditional perturbation measures.

For strain, acoustic measures are most useful as part of a profile that includes maximum phonation time, S/Z ratio, and connected-speech observations. A single number is unlikely to do justice to the dimension.

Multiparametric indices like AVQI, ABI, and CSID combine several acoustic parameters into a single score that correlates well with overall dysphonia severity. They are valuable: a global severity number is faster to interpret, easier to track over time, and more robust than any single parameter.

But severity is not the same as roughness, and roughness is not the same as breathiness or strain. The parameter ranking changes by dimension in ways that matter clinically:

A patient with mild breathiness from incomplete closure and a patient with mild strain from hyperfunction can produce the same AVQI score for very different reasons. The dimensions provide clinically actionable information that the global score collapses. This is not a criticism of multiparametric indices—it is an argument for using them alongside dimension-specific measures, not in place of them.

For a clinician who already uses CAPE-V or GRBAS perceptually, acoustic analysis adds value in three concrete ways:

A complete picture of voice quality requires more than acoustic measurement. The CAPE-V protocol exists because voice quality involves perception that integrates context, connected speech, phonatory onset, and registration in ways that sustained vowel acoustics do not capture. Several clinically important phenomena are systematically under-detected by acoustic measures:

• Onset and offset patterns—glottal attack, breathy attack, and hard offsets
• Connected-speech instability—phonation breaks, register transitions, vocal fry segments
• Pitch breaks and diplophonia—often present in connected speech but not in sustained vowels
• Vocal effort that does not produce measurable signal disturbance—pressed voice with clean periodicity
• Resonance abnormalities—hypernasality, cul-de-sac resonance, that affect perceived voice quality without primarily originating at the larynx

Perceptual judgment by a trained clinician integrates these in a way that no current acoustic system does. The role of acoustic analysis is to make the perceptual judgment more reproducible and easier to document, not to replace it.