13 Acoustic Voice Quality Index (AVQI)
Living draft. This chapter reflects the sources cited below and will be revised as further sources are incorporated. Found an error? Use the Report an issue link in the sidebar.
13.1 Definition
The question the AVQI asks is how severe the voice sounds overall. Unlike the measures in Parts II and III it asks nothing of the waveform directly: it combines several of them into one score aimed at perceived overall quality.
The Acoustic Voice Quality Index (AVQI) is an objective method to quantify the severity of overall voice quality in concatenated continuous speech and sustained phonation segments (Barsties and Maryn 2016). It is a multivariate construct based on linear regression that combines six acoustic markers into a single score for the whole voice sample (Barsties and Maryn 2016). That score represents the overall voice quality — the hoarseness level — of the sample (Delgado Hernández et al. 2018).
The index was proposed by Maryn, Corthals, et al. (2010) and was one of the first measures to combine both speech types (Delgado Hernández et al. 2018). The rationale for the combination: sustained vowels are easily elicited and less affected by articulation and dialect, while continuous speech is more representative of a patient’s daily voice (Delgado Hernández et al. 2018).
13.2 Computation
13.2.1 Speech material and concatenation
The subject sustains the vowel [a:] for longer than 3 seconds, and a 3-second mid-vowel portion is selected. The subject also reads a phonetically balanced text aloud at comfortable pitch and loudness. For AVQI version 03.01 in Dutch, the continuous-speech part is trimmed to the first 34 syllables of the text, the 3-second [a:] segment is appended, and the result is saved as a single sound file (Barsties and Maryn 2016). Acoustic analysis is applied only to the voiced segments of the continuous speech — extracted by an automated Praat detection script — plus the appended vowel segment (Barsties and Maryn 2016).
13.2.2 Component measures
All six parameters are obtained in Praat from the concatenated signal (Barsties and Maryn 2016; Delgado Hernández et al. 2018):
| Parameter | Definition |
|---|---|
| Smoothed cepstral peak prominence (CPPS) | Distance between the first rahmonic’s peak and the point of equal quefrency on the regression line through the smoothed cepstrum — see the CPPS chapter |
| Harmonics-to-noise ratio (HNR) | Base-10 logarithm of the ratio between periodic energy and noise energy, × 10 |
| Shimmer local (Shim) | Absolute mean difference between amplitudes of successive periods, divided by the average amplitude |
| Shimmer local dB (ShdB) | Base-10 logarithm of the difference between amplitudes of successive periods, × 20 |
| General slope of the spectrum (Slope) | Difference between the energy in 0–1000 Hz and in 1000–10000 Hz of the long-term average spectrum |
| Tilt of the spectral trendline (Tilt) | The same difference computed on the trendline through the long-term average spectrum |
13.2.3 Regression equation (version 03.01)
As printed by Barsties and Maryn (2016):
\[ \text{AVQI}_{03.01} = \bigl(4.152 - 0.177\,\text{CPPS} - 0.006\,\text{HNR} - 0.037\,\text{Shim} + 0.941\,\text{ShdB} + 0.01\,\text{Slope} + 0.093\,\text{Tilt}\bigr) \times 2.8902 \]
The published Praat script displays the resulting score on a 0–10 axis, marked green below the Dutch threshold of 2.43 and red above it (Barsties and Maryn 2016).
Two validation papers print this equation differently. Barsties v. Latoszek et al. (2020) print the HNR term as “(0.0006-HNR)” and Delgado Hernández et al. (2018) as “(0.0006 − HNR)”, both with a minus sign where a multiplication is expected; Delgado Hernández et al. (2018) also print the Tilt coefficient as 0.0993. The other sources — Barsties and Maryn (2016) (text and appendix script), Englert et al. (2021), and Castillo-Allendes et al. (2023) — give 0.006 × HNR and 0.093 × Tilt, as above. All printed forms were checked against the publisher PDFs, so the differences are in the publications. When implementing the index, take the coefficients from the Praat script printed in the appendix of Barsties and Maryn (2016) rather than from an equation retyped in a later paper.
A third discrepancy is of another kind. Aghajanzadeh et al. (2025) name AVQI 03.01 throughout and computed it in VOXplot, but the equation they print is the development study’s, coefficient for coefficient, with its ×2.571 rescaling (Maryn, Corthals, et al. 2010). Either the printed equation or the version label is wrong, and the paper does not let a reader tell which. The Persian threshold below stays in the table, with this caveat attached.
13.2.4 Versions 02.02 and 03.01
In the earlier AVQI versions the analyzed continuous-speech part was much shorter than the constant 3-second vowel part (1.53 s of voiced speech in Dutch, 1.99 s in Japanese) (Delgado Hernández et al. 2018). Version 03.01 balances the analyzed durations of the two speech types by expanding the continuous-speech sample from 17–22 syllables to around 34 syllables, for higher ecological validity and balanced internal consistency (Barsties and Maryn 2016; Delgado Hernández et al. 2018). Early AVQI workflows required both SpeechTool and Praat; once CPPS was implemented in Praat, the index ran in Praat alone, with highly comparable outcomes (Barsties and Maryn 2016).
The regression-based combination of measures over a concatenated sample was explored for tracheoesophageal (alaryngeal) voice by Maryn et al. (2009), where a two-factor model combining CPP with the prominence of the second spectral harmonic reached rs = 0.87 with listener judgments.
13.3 What it captures
The AVQI quantifies overall voice quality as judged by the grade (G) of the GRBAS scale — the degree of hoarseness or voice abnormality (Barsties and Maryn 2016; Delgado Hernández et al. 2018). Concurrent validity with mean perceptual G is strong: r = 0.815 in the Dutch external validation (66.4% of variance) (Barsties and Maryn 2016) and r = 0.835 in Spanish (69.7% of variance) (Delgado Hernández et al. 2018). CPPS is the main factor in the multivariate model (Barsties and Maryn 2016). Barsties v. Latoszek et al. (2018) report, from the earlier overall-voice-quality meta-analysis of Maryn, Roy et al. (2009), weighted correlations for CPPS of 0.63 for sustained vowels and 0.88 for continuous speech.
A previous AVQI investigation (Maryn, De Bodt, et al. 2010) reported high sensitivity to voice changes through voice therapy (Barsties and Maryn 2016).
13.4 Normative data
Diagnostic thresholds separate normophonic from hoarse voices, with hoarseness defined as mean perceptual grade G ≥ 0.5 (Barsties and Maryn 2016). The threshold differs between AVQI versions and languages (Barsties v. Latoszek et al. 2020), so each row states its language and pipeline. Use the row validated for the patient’s language and the AVQI version in use; see Pipeline Dependence and Reference Values.
| Language | Sample | Pipeline | Threshold | Sensitivity | Specificity | AUC | Class |
|---|---|---|---|---|---|---|---|
| Dutch, external validation | 970 dysphonic, 88 normal | AVQI 03.01; Praat 5.3.57 | 2.43 | 0.785 | 0.932 | 0.923 | language-conditional |
| Spanish | 136 dysphonic, 47 normal | AVQI 03.01; Praat 6.0.22 | 2.28 | 0.748 | 0.946 | 0.905 | language-conditional |
| German | 175 dysphonic, 43 healthy; 27 syllables | AVQI 03.01; Praat 5.3.57 | 1.85 | 0.72 | 0.90 | 0.90 | language-conditional |
| French | 90 ENT patients, 30 controls; 27 syllables | AVQI 03.01; Praat version not stated; recorded at 11,025 Hz | 2.33 | 0.598 | 1.00 | 0.788 | language-conditional |
| Brazilian Portuguese | 113 dysphonic, 37 nondysphonic; counting 1–11 (17 syllables) | AVQI 03.01 script; Praat 6.0.40 | 1.33 | 0.788 | 0.906 | 0.904 | language-conditional |
| Brazilian Portuguese, counting 1–11 | 40 dysphonic, 13 healthy; 17 syllables | AVQI 03.01; Praat 6.0.06 | 1.16 | 0.86 | 0.80 | 0.87 | language-conditional |
| Brazilian Portuguese, reading text | same 53 speakers; 32 syllables | AVQI 03.01; Praat 6.0.06 | 1.56 | 0.886 | 1.00 | 0.963 | language-conditional |
| US English | 148 with voice disorders, 49 non-clinic-seeking; first 22 syllables of the Rainbow Passage | AVQI 03.01 in VOXplot 2.0.0; Praat 6.3.06 for extraction | 1.17 | 0.62 | 0.95 | 0.84 | language-conditional |
| Japanese | 455 recordings, 55 from controls; 30 syllables, time-balanced | AVQI 03.01; Praat version not stated | 2.06 | 0.721 | 0.938 | 0.915 | language-conditional |
| Persian | 136 dysphonic, 45 vocally healthy; 20 syllables, durations balanced | AVQI 03.01 in VOXplot, version not stated | 1.63 | 0.732 | 0.91 | 0.850 | language-conditional |
| American English, PVQD; candidate, exploratory | 181 with voice disorders, 109 vocally healthy; 288 rated; sustained /a/ plus concatenated CAPE-V sentences; GRBAS grade ≥ 0.5 | AVQI 03.01 in PhonaLab; Parselmouth 0.4.7 (Praat 6.1.38 engine) | 1.96 | 0.783 | 0.753 | 0.838 | language-conditional |
| American English, PVQD; candidate, exploratory | the same sample and tasks; CAPE-V overall severity ≥ 10 | the same pipeline | 1.96 | 0.734 | 0.784 | 0.825 | language-conditional |
| Pooled: eight AVQI v03 studies (context only) | 12 languages | multiple | none; study thresholds ranged 1.33–3.15 | 0.82 | 0.92 | 0.94 (SROC) | meta-analytic |
The pooled row is descriptive context, not a cutoff. Its sensitivity was highly heterogeneous across studies (I² = 95.3%), its specificity much less so (I² = 16.4%) (Jayakumar and Benoy 2024).
The two American English PVQD rows are candidate cutoffs, not validated ones. The analysis that produced them was a secondary, exploratory objective, each cutoff was chosen by Youden’s J index on the full sample, and the study states that they require external validation before clinical use (Lucero 2026). Ten-fold cross-validation stayed close to the resubstitution figures, with AUC optimism of 0.002 or less. The author of that study developed and operates PhonaLab.
They do not disagree with the US English row above them. That row computes AVQI 03.01 in VOXplot 2.0.0 from 22 syllables of the Rainbow Passage and classifies voices by mean grade (Castillo-Allendes et al. 2023); these rows compute it in PhonaLab, through Parselmouth 0.4.7, from concatenated CAPE-V sentences, and classify voices by CAPE-V overall severity or GRBAS grade (Lucero 2026). Different pipeline, different speech material, different reference standard, so the rows are not comparable and the distance between their thresholds is not a disagreement between studies. That both reference standards return the same candidate cutoff is a result about this sample, not a cross-validation of it.
One further set of figures belongs to the Dutch internal validation, and is not published as a second Dutch cutoff. Barsties and Maryn (2015), the 60-sample study that produced the version 03.01 weighting, reports for its balanced-duration condition an AROC of 0.98 at a threshold of 2.43, with sensitivity 1.00 and specificity 0.936. These are development-sample figures: the threshold was evaluated on the same recordings that produced the model’s weights, which is what a sensitivity of exactly 1.00 and an infinite positive likelihood ratio indicate. The Dutch row of Table 13.2 therefore carries the external validation on 1058 samples instead (Barsties and Maryn 2016). Note that the secondary literature transposes the two figures, giving sensitivity 0.936 and specificity 1.00.
The Japanese threshold needs care for a different reason, set out under Threshold derivation criteria below. A separate paper by the same group validated AVQI version 02.02 in Japanese and reports a threshold of 3.15, with sensitivity 72.5% and specificity 95.2%, on a standardized syllable number of 22 (Hosokawa et al. 2017). That is not a rival estimate of the 2.06 in the table: the two belong to different versions of the index, with different regression weights and different speech-material lengths, so 3.15 cannot be compared with any threshold in Table 13.2. Always pair an AVQI threshold with its version.
Threshold derivation criteria. The thresholds in Table 13.2 are not all derived the same way. The Japanese validation computed candidates from both the Youden Index and the upper 95th percentile of its control cohort. The Youden Index gave 1.41, with balanced sensitivity 84.4% and specificity 85.6%; the 95th percentile gave 2.06, at 72.1% and 93.8%; and the original Dutch threshold of 2.43 applied to the same sample gave 64.6% and 97.3% (Hosokawa et al. 2019). The authors recommend 2.06, because at that value sensitivity and specificity came close to what the Dutch threshold produced in their cohort, and the table follows them (Hosokawa et al. 2019). The consequence is that the Japanese row trades sensitivity for specificity relative to a Youden-derived cutoff, so comparing it with the Youden-derived rows compares two criteria and not only two languages. Some of the spread in published AVQI thresholds is a spread of derivation methods rather than of languages.
Likelihood ratios at the optimal threshold: LR+ = 11.54, LR− = 0.23 (Dutch external validation) (Barsties and Maryn 2016); LR+ = 13.85, LR− = 0.26 (Spanish, threshold chosen by the Youden index) (Delgado Hernández et al. 2018); LR+ = 7.40, LR− = 0.31 (German) (Barsties v. Latoszek et al. 2020); LR+ = 8.38, LR− = 0.23 (Brazilian Portuguese) (Englert et al. 2021); LR+ = 12.46, LR− = 0.40 (US English) (Castillo-Allendes et al. 2023). The LR+ values printed by Englert et al. (2020) are below 1 or negative and cannot be correct, so they are not reproduced here.
The original AVQI, which carried no version number, reported two cutoff options: 2.36 (sensitivity 91%, specificity 59%) and 2.95 (sensitivity 74%, specificity 96%). The authors suggest 2.95 may be more appropriate when the main interest is screening (Maryn, Corthals, et al. 2010). These values apply to that version only.
The US English validation standardized the continuous-speech part at 22 syllables of the Rainbow Passage, recorded in a quiet environment (ambient noise below 35 dB) with a head-mounted microphone at 44.1 kHz / 16 bit, and computed the index in VOXplot 2.0.0, which the authors state uses established Praat algorithms and the same equation (Castillo-Allendes et al. 2023). Its threshold, 1.17, lies below the range of the AVQI meta-analysis, whose literature search ended in July 2021 (Jayakumar and Benoy 2024). The authors note that other English dialects and accents remain to be examined (Castillo-Allendes et al. 2023).
13.5 Confounds and cautions
Language-specific standardization is required. The continuous-speech part must be trimmed to a syllable count that yields about 3 seconds of voiced speech, and the count differs by language: 34 syllables in Dutch, 33 in Spanish, 30 in Japanese for AVQIv3 (Delgado Hernández et al. 2018).
Thresholds differ by language, by version, and by derivation criterion. Reported optimal AVQI version 03.01 cutoffs are 2.43 (Dutch), 2.28 (Spanish) and 2.06 (Japanese). The differences are small, which may indicate that AVQI is relatively insulated from inter-language phonetic differences (Delgado Hernández et al. 2018). The validations since widen the spread: 1.85 German, 2.33 French, 1.33 Brazilian Portuguese, 1.63 Persian. A version 02.02 threshold sits on a different scale altogether, as the Japanese version 02.02 value of 3.15 shows (Hosokawa et al. 2017). Nor is a percentile-derived threshold directly comparable with a Youden-derived one, which is the case for the Japanese row.
Equal proportion of the two speech tasks. Earlier work found the sustained vowel dominating the final AVQI score more than the continuous speech. Barsties and Maryn (2015) tested three speech-material durations against that: 17 syllables plus 3 s of sustained phonation, a customised continuous-speech length plus 3 s, and a whole text plus 3 s. Concurrent validity was comparable across all three (rs 0.891 to 0.929), but precision was best, and the two tasks’ contributions balanced, in the middle condition. This is the reasoning behind the standardized syllable count that the later validations use.
Recording conditions matter. Validation recordings were made in sound- treated rooms with head-mounted condenser microphones at 44.1 kHz / 16 bit, with environmental noise verified against a signal-to-noise norm (Barsties and Maryn 2016; Delgado Hernández et al. 2018). The AVQI script itself states that reliable analysis requires optimal data-acquisition conditions (Barsties and Maryn 2016).
The Praat version matters. With the same AVQI 03.01 script and the same 218 German recordings, seven Praat versions formed three clusters. Praat 6.0.44 introduced a bug in the CPPS computation that 6.0.47 fixed. With version 6.0.46, AVQI scores were on average 4.16 points lower, and sensitivity at the German threshold of 1.85 fell from 72% to 23%. Across the unaffected versions, mean differences were negligible, but single recordings differed by up to 0.62 points (Stappenbeck et al. 2020). The authors recommend using the Praat version of the validation study for the language in question, or checking a new version against an earlier one on test recordings (Stappenbeck et al. 2020).
Sampling rate. The French validation recorded at 11,025 Hz (Pommée et al. 2020), so its signals contain nothing above about 5.5 kHz. The AVQI script computes the LTAS slope and trend-line tilt over bands that reach 10 kHz (Barsties and Maryn 2016), so two of the six components are computed on a narrower band than in the other validations. Treat the French threshold as applying to recordings made at that rate.
Report the version. Versions 02.02 and 03.01 differ in the length of the continuous-speech part (17 vs. 34 syllables) and in the regression weighting (Barsties and Maryn 2016). Report the version with every score.
Base rates affect ROC statistics. The Dutch external validation sample contained 8% normophonia and 92% dysphonia; likelihood ratios were used because they are less vulnerable to base-rate differences than ROC statistics alone (Barsties and Maryn 2016).
Voicing detection is part of the pipeline. Voiced segments of continuous speech are extracted automatically by a Praat script before measurement (Barsties and Maryn 2016).
13.6 Validation
Across 1118 subjects in two studies, AVQI 03.01 showed a homogeneous weighted mean correlation of r = 0.821 with perceptual overall voice quality, against r = 0.790 for the initial AVQI model pooled over 507 subjects (Barsties and Maryn 2016). In the same pooled analysis, the Dysphonia Severity Index reached a heterogeneous weighted mean r = 0.524 and the Cepstral Spectral Index of Dysphonia r = 0.788 (sustained vowel) and 0.748 (continuous speech); AVQI 03.01 reached the highest concurrent validity among these multiparametric indices (Barsties and Maryn 2016). Language validations include Dutch (Barsties and Maryn 2016) and Spanish (Delgado Hernández et al. 2018), with Japanese figures reported in the latter.
13.7 Compute it in PhonaLab
PhonaLab is developed by the author of this reference; see Competing interests.
Upload a sustained vowel and a continuous speech sample to compute AVQI on your own recording, with the version and language protocol reported alongside the score.