5 Perceptual Anchors (GRBAS, CAPE-V)
Living draft. This chapter explains the perceptual instruments that every diagnostic threshold in this reference was derived against. The composite chapters point here from their Normative data sections. Found an error? Use the Report an issue link in the sidebar.
5.1 Why a chapter on something that is not a measure
GRBAS and the CAPE-V are auditory-perceptual rating instruments. A listener hears a voice and scores it on named dimensions. Neither is computed from a signal, and neither is an acoustic measure.
They are here because every composite index in Part IV was fitted against them, and because every diagnostic threshold in this reference was obtained by asking how well an acoustic value separates two groups that a perceptual instrument defined. A cutoff is a statement about a perceptual standard as much as about a measure. Change the instrument, change the dimension, or move the point on the scale at which a voice counts as affected, and the cutoff moves without anything about the acoustics changing.
That is the argument of Pipeline Dependence and Reference Values, one layer up. A measured value belongs to the pipeline that produced it; a cutoff belongs to the pipeline and to the perceptual standard that defined its groups.
The ASHA expert panel treats auditory-perceptual assessment with the CAPE-V as a separate, non-instrumental part of a voice evaluation, accompanying acoustic, laryngeal-imaging and aerodynamic assessment rather than being replaced by them (Patel et al. 2018). Nothing in this reference argues otherwise. An acoustic index approximates a perceptual judgement; it does not supersede it.
5.2 GRBAS
The scale was proposed by the Committee for Phonatory Function Tests of the Japan Society of Logopedics and Phoniatrics, for evaluating hoarseness. It consists of five scales (Hirano 1981).
| Scale | What it represents | Stated acoustic correlate |
|---|---|---|
| G, grade | The degree of hoarseness or voice abnormality; it corresponds to the factor of evaluative nature obtained by the semantic differential technique | — |
| R, rough | The psycho-acoustic impression of irregularity of vocal fold vibrations | Irregular fluctuations in the fundamental frequency and/or the amplitude of the glottal source sound |
| B, breathy | The psycho-acoustic impression of the extent of air leakage through the glottis | Related to turbulence |
| A, asthenic | Weakness or lack of power in the voice | A weak intensity of the glottal source sound and/or a lack of higher harmonics |
| S, strained | The psycho-acoustic impression of a hyperfunctional state of phonation | An abnormally high fundamental frequency, noise in the high frequency range, and/or richness in high frequency harmonics |
Each scale is graded on four points: 0 non-hoarse or normal, 1 slight, 2 moderate, 3 extreme (Hirano 1981). A result is written as the five scores together, so a GRBAS rating is a profile rather than a number.
Two cautions come from the proposing committee itself. The evaluation is subjective, so the examiner must possess a trained ear, and a standard tape of typical voice samples was made for training. And the committee states that psycho-acoustic evaluation with the GRBAS scale is not an absolute method and needs to be improved upon (Hirano 1981). Both remarks are from 1981 and neither has been overtaken.
5.3 CAPE-V
The CAPE-V was developed under ASHA Special Interest Division 3, first released in 2002 and revised in 2009 (Patel et al. 2018). It is scored on a visual analog scale, and its dimensions include overall severity, roughness, breathiness, strain, pitch and loudness (Walden 2022). Its sentences, recorded for the auditory-perceptual assessment of voice quality, can also serve as the connected-speech sample for cepstral analysis (Patel et al. 2018).
This reference does not yet describe the CAPE-V from its own documentation. The 2009 revision (Kempster and colleagues, as cited by Patel et al. (2018)) is the instrument’s source of record for its tasks, its visual anchors and its scoring, and it is not yet in this project’s source library. Until it is, this chapter states only what other sources report about the instrument. That is a declared absence of a source, in the same spirit as a declared absence of a reference value: see Pipeline Dependence.
5.4 Averaging, and the thresholds it creates
Both instruments are applied by individual listeners. Every validation study in this reference uses a mean across raters as its reference value.
That has a consequence worth stating plainly. Averaging turns GRBAS from a four-point ordinal scale into a continuum, which is why thresholds such as a mean grade of 0.5 exist at all. No single rater can return 0.5 on a scale whose points are 0, 1, 2 and 3. The threshold is a property of the panel, not of a judgement any listener made.
5.5 The reference standards behind this reference’s cutoffs
Table 5.2 is the set of perceptual standards that the cutoff studies cited in Part III and Part IV actually used. The column that matters is the third one.
| Study and language | Instrument | A voice counted as affected when | Raters and reliability |
|---|---|---|---|
| Dutch, AVQI external validation | GRBAS grade, mean | \(G_\text{mean} \geq 0.5\) | — |
| Brazilian Portuguese, AVQI and ABI | GRBAS grade and breathiness, means | G < 0.5 and B < 0.5 taken as unaffected | five judges; Fleiss \(\kappa\) 0.370 for G, 0.353 for B |
| Finnish, ABI | GRBAS breathiness, mean | a nonbreathy voice was one rated B 0–0.49 | eight listeners; Fleiss \(\kappa\) 0.435 |
| Malayalam and Kannada, ABI | GRBAS breathiness | “B” of GRBAS as the reference standard | — |
| US English, AVQI and ABI | GRBAS grade, mean | \(G_\text{mean} \geq 0.5\), stated for both indices | — |
| American English, PVQD, AVQI and CPPS | CAPE-V overall severity, 100-point scale | severity \(\geq\) 10 | nineteen raters, three per file |
| American English, PVQD, AVQI | GRBAS grade, mean | grade \(\geq\) 0.5 | as above |
| American English, PVQD, ABI | CAPE-V breathiness, 100-point scale | breathiness \(\geq\) 10 | as above |
| Six ABI validations, pooled | GRBAS, means | per study | three to five judges per study; Fleiss \(\kappa\) 0.353 to 0.492 |
Two of those choices carry a stated rationale. On the 100-point scale, 10 was selected as a clinically meaningful boundary between normal variation and minimal perceptible deviance; on the averaged ordinal scale, 0.5 is the midpoint between normal and minimal severity (Lucero 2026).
The ABI meta-analysis records a reliability floor for this literature: an inter-rater Fleiss \(\kappa\) of 0.41 is the recommended minimum, and the one included study that fell below it is the one whose correlation between ABI and perceived breathiness was lowest (Barsties v. Latoszek et al. 2021).
5.6 The PVQD, in detail
The Perceptual Voice Qualities Database is the sample behind this reference’s American English rows, so its rating procedure deserves its own section.
Its 296 recordings were rated by nineteen experienced voice clinicians using voice qualities from the CAPE-V, without visual anchors, and the GRBAS scale (Walden 2022). Raters had to be voice professionals with at least two years of experience, familiar with both scales and using at least one of them monthly; twenty-five were recruited and nineteen completed (Walden 2022).
| Scale | Inter-rater ICC | Intra-rater ICC |
|---|---|---|
| 100-point visual analog | 0.86 | 0.913 |
| GRBAS | 0.859 | 0.889 |
Three cautions apply to those figures, and all three come from the database paper itself (Walden 2022).
They are reliability for a mean of several raters. The ICCs assume final ratings based on three raters for inter-rater reliability and two repeated ratings for intra-rater reliability, and were corrected upward with the Spearman-Brown formula to account for the higher number of ratings per target. A single clinician rating a voice alone is not at 0.918.
They are consistency, not agreement. Because different raters rated different subsets of the database, reliability can only be assessed as consistency between or within raters, under Shrout and Fleiss’s ICC(1, m) model. The design cannot assess agreement in the levels of the ratings. The same non-overlapping assignment is why the validation study built on this database could not model individual rater effects and used the published means instead (Lucero 2026).
The severity scale is CAPE-V-like, not the CAPE-V. Its dimensions were borrowed from the CAPE-V and scored on a 100-point visual analog scale without the instrument’s visual anchors. A threshold written as “CAPE-V overall severity \(\geq\) 10” on this database is a threshold on that scale.
One further point concerns what “vocally healthy” means here. The database paper reports 296 speakers, of whom 89 had no voice complaint, 187 reported a complaint or had a confirmed diagnosis of dysphonia, and 20 have no data recorded for voice complaint or diagnosis (Walden 2022). The validation study using it reports 290 usable speakers as 181 with voice disorders and 109 vocally healthy (Lucero 2026). The arithmetic reconciles only if the 20 speakers with no recorded voice status are counted among the 109 controls, and all six exclusions came from the disordered group. On that reading, roughly one in five of the controls behind this reference’s American English specificity figures has no documented voice status. The inference is from the counts in the two papers, and it should be confirmed against the database labels before it is relied on.
5.7 How to read a threshold
- A threshold inherits its reference standard. A cutoff validated against mean GRBAS grade answers the question “would a panel of listeners have called this voice hoarse?”. A cutoff validated against CAPE-V breathiness answers a different question about a different dimension. A measure may separate one well and the other poorly, which is the point of a dimension-targeted index: see ABI and GNE.
- Two reference standards on one sample are not two validations. Where a study reports the same cutoff against both CAPE-V severity and GRBAS grade, that is one sample analysed twice. See AVQI.
- A dichotomised perceptual rating is a summary, not a diagnosis. The instruments grade a voice. They do not identify a lesion or a mechanism (Patel et al. 2018).
- Report the instrument, the dimension and the threshold whenever a cutoff is quoted, exactly as the pipeline is reported for the measure itself.
5.8 Where the sources disagree
Where “normal” ends on an averaged GRBAS grade. The database paper’s own figure categories place the normal band at 0 to 0.5, mild at 0.6 to 1.5, moderate at 1.6 to 2.5 and severe at 2.6 to 3.0 (Walden 2022). The validation literature treats a mean grade of 0.5 or above as affected (Barsties and Maryn 2016). A voice with a mean grade of exactly 0.5 is therefore normal under the first convention and dysphonic under the second. No value in this reference is moved on account of it. The point matters because the boundary sits where the ratings are densest: the PVQD’s severity distribution is skewed toward the normal, mild and moderate levels, and perception of mild-to-moderate severities is harder than of normal and severe ones (Walden 2022).
Whether rating scales are reliable enough to serve as a reference standard at all. The database paper establishes good inter- and intra-rater reliability for the PVQD overall and concludes that its ratings may be used with relative confidence for research (Walden 2022). A standard textbook treatment reaches a much darker verdict about dimensional rating protocols as measurement tools: across studies, scales and statistics, average interrater reliability has ranged from extremely low (\(r^2 = .04\)) to extremely high; listeners almost never agree in their ratings of a single voice; the likelihood that two raters agree on moderately pathological voices on seven-point scales averaged 0.21 where chance is 0.14; and more than 60%, and as much as 78%, of the variance in voice quality ratings was attributable to factors other than differences among voices in the quality being rated (Kreiman and Sidtis 2011).
Both can hold at once, and the difference is in what is being measured. The ICCs are for a mean of three trained raters on one database. The textbook figures are for individual listeners judging individual voices across the literature. The practical consequence is the same either way, and it is why this chapter exists: the perceptual threshold is the softest element in any acoustic cutoff, and it should be reported rather than assumed.