4  Pipeline Dependence and Reference Values

Note

Living draft. This chapter explains why every reference value in this reference is reported with the pipeline that produced it. Measure chapters point here from their Normative data sections. Found an error? Use the Report an issue link in the sidebar.

4.1 A value belongs to its pipeline

An acoustic voice measure is produced by a pipeline: the recording chain, the analysis program and its version, the algorithm and its settings, the speech material, and, for composite indices, the language and the perceptual reference used to set a cutoff. The number that comes out depends on the pipeline as well as on the voice. A cutoff or a normative value is therefore valid only for measurements made with the same or a compatible pipeline.

This is easy to miss. Programs label their output with the same names, such as “mean jitter” or “percent shimmer”, and the shared labels suggest that results from different programs are comparable (Bielamowicz et al. 1996). This chapter shows, from published comparisons, how far that suggestion fails, and sets out how the reference reports reference values as a result.

4.2 What counts as the pipeline

Table 4.1: Pipeline elements shown to change measured values.

Pipeline element	Evidence that it changes the value
Analysis program	Three systems agreed on mean F0 (within 3 Hz) but not on jitter or shimmer, even with standardized recording and analysis procedures (Karnell et al. 1995).
Implementation of a named measure	Two programs that both compute relative average perturbation (RAP) gave significantly different percent jitter (Bielamowicz et al. 1996).
Algorithm details	For CPP, ADSV and Praat differ in voicing detection, window, interpolation, regression fit and regression range (Watts et al. 2017).
Program settings	The same program gave a Rainbow Passage CPPS cutoff of 19.10 dB with factory settings (Sauder et al. 2017) and 9.33 dB with published non-default settings (Murton et al. 2020).
Program version	Three CPPS studies used three Praat versions (6.0.17, 6.0.40, 6.0.50), and the version should be reported (Buckley et al. 2023). One Praat release changed AVQI and ABI scores enough to invalidate their cutoffs (Stappenbeck et al. 2020).
Speech material	Stimulus had a large effect on cepstral values, larger than software; sex had a small effect (Buckley et al. 2023).
Portion of a passage	Studies analysed the second sentence, the first 12 s, or sentences 2–3 of the Rainbow Passage (Sauder et al. 2017; Murton et al. 2020; Buckley et al. 2023).
Index version and language	The AVQI threshold differs between versions and languages (Barsties v. Latoszek et al. 2020).
Perceptual reference	The AVQI development study counted a voice as normal only if all judges rated G = 0 (Maryn et al. 2010); later validations used mean G below 0.5 (Englert et al. 2021; Pommée et al. 2020).
Pre-processing	Digital differencing combined with a hardware pre-emphasis filter pre-emphasizes the signal twice (Kent and Read 2002, 64).
Automatic segment rejection	A program must not automatically and surreptitiously discard segments it deems unacceptable (Baken and Orlikoff 2000, 132).

4.3 Evidence 1: perturbation measures across programs

The earliest systematic comparisons concerned jitter and shimmer. Karnell and colleagues analysed the same 60 vowel tokens from 20 patients with three systems. Mean F0 differed by less than 3 Hz, with correlations of .98 to 1.00. Jitter correlations between systems were .29, .62 and .64, and shimmer correlations ranged from .26 to .42 (Karnell et al. 1995). The authors advise clinicians who use perturbation measures as objective measures to examine their validity and reliability carefully (Karnell et al. 1995).

Bielamowicz and colleagues compared four systems on 50 dysphonic voices. On the near-periodic (type 1) signals that all systems could analyse, rank-order correlations ranged from .33 to .80 for jitter and from .81 to .89 for percent shimmer, and mean HNR/SNR on the same voices ranged from 8.58 to 21.31 dB across systems (Bielamowicz et al. 1996). They note that a high rank-order correlation may mask significant differences in the values themselves, and that programs which hide where cycle boundaries were placed make it virtually impossible to check whether a perturbation analysis is valid (Bielamowicz et al. 1996).

4.4 Evidence 2: cepstral peak prominence across programs

Cepstral peak prominence is often recommended because it does not need F0 tracking (see Signal Typing). It is not pipeline-free. The two programs most used clinically, ADSV and Praat, compute it differently. ADSV discards frames with CPP below 0 dB as a simple voicing detector, fits the regression line by least squares from 0.0001 s quefrency, and uses a Hamming window without interpolation. Praat applies no voicing detection, fits a robust (Theil) line from 0.001 s, and uses a Gaussian window with parabolic interpolation (Watts et al. 2017).

The consequences are large:

• On the same MEEI recordings of sustained /a/, the cutoff separating voice-disordered from typical speakers was 11.46 dB for ADSV CPP and 14.45 dB for Praat CPPS (Murton et al. 2020).
• On the same 170 Rainbow Passage samples, ADSV CPPS ranged from 0.40 to 8.31 dB and Praat CPPS from 14.71 to 22.99 dB: the two ranges do not overlap. The programs correlated at r = 0.88, and the authors conclude that absolute values from the two programs cannot be compared directly (Sauder et al. 2017).
• Watts and colleagues found Praat and ADSV CPP correlated at r = 0.92 to 0.96 in English, yet clearly different in magnitude. They suggest the difference may be due particularly to the regression-line calculation, and recommend comparing CPP data only with data from the identical program and algorithm (Watts et al. 2017).

4.5 Evidence 3: one program, different settings and material

Changing the settings of a single program also moves the reference. With Praat 6.0.17 at factory settings on the second sentence of the Rainbow Passage, the CPPS cutoff was 19.10 dB (Sauder et al. 2017). With the settings published by Watts and colleagues, on the first 12 s of the passage, the Praat CPPS cutoff was 9.33 dB (Murton et al. 2020). Buckley and colleagues point out both differences: the settings, and the portion of the passage analysed (Buckley et al. 2023).

Buckley and colleagues then measured 150 speakers without voice disorders, aged 18 to 91 years, with the same Praat settings as Murton and colleagues. 56% of the women and 65.3% of the men fell below the 9.33-dB cutoff, and all fell below the 19.10-dB cutoff. They conclude that cutoffs from ROC analysis do not adequately represent speakers without voice disorders, and propose normative lower limits 2 SD below the group mean instead: 6.40 dB (men) and 6.49 dB (women) for the Rainbow Passage in Praat (Buckley et al. 2023).

Figure 4.1 places these values on one axis.

# Values as published. Kind: "cut" = ROC cutoff, "norm" = normative lower limit.
rainbow = [
    ("ADSV", "Murton 2020 (first 12 s)", 6.11, "cut"),
    ("ADSV", "Sauder 2017 (sentence 2, factory)", 5.53, "cut"),
    ("ADSV", "Buckley 2023, men (sentences 2–3)", 5.40, "norm"),
    ("ADSV", "Buckley 2023, women (sentences 2–3)", 5.04, "norm"),
    ("Praat", "Murton 2020 (first 12 s, Watts settings)", 9.33, "cut"),
    ("Praat", "Sauder 2017 (sentence 2, factory)", 19.10, "cut"),
    ("Praat", "Buckley 2023, men (Watts settings)", 6.40, "norm"),
    ("Praat", "Buckley 2023, women (Watts settings)", 6.49, "norm"),
]
vowel = [
    ("ADSV", "Murton 2020", 11.46, "cut"),
    ("ADSV", "Buckley 2023, men", 8.86, "norm"),
    ("ADSV", "Buckley 2023, women", 8.09, "norm"),
    ("Praat", "Murton 2020 (Watts settings)", 14.45, "cut"),
    ("Praat", "Buckley 2023, men (Watts settings)", 11.72, "norm"),
    ("Praat", "Buckley 2023, women (Watts settings)", 11.05, "norm"),
]
colors = {"ADSV": vs.ORANGE, "Praat": vs.BLUE}

fig, axes = plt.subplots(2, 1, figsize=(7.2, 5.8), sharex=True,
                         gridspec_kw={"height_ratios": [len(rainbow), len(vowel)]})
for ax, data, title in zip(axes, [rainbow, vowel], ["Rainbow Passage", "Sustained /a/"]):
    y = np.arange(len(data))[::-1]
    for yi, (prog, label, val, kind) in zip(y, data):
        c = colors[prog]
        if kind == "cut":
            ax.plot(val, yi, marker="D", ms=7, color=c, zorder=3)
        else:
            ax.plot(val, yi, marker="o", ms=7, mfc="white", mec=c, mew=1.8, zorder=3)
        ax.text(val + 0.35, yi, f"{val:.2f}", va="center", fontsize=8, color=vs.INK)
    ax.set_yticks(y)
    ax.set_yticklabels([f"{p}: {l}" for p, l, _, _ in data], fontsize=8)
    ax.set_ylim(-0.7, len(data) - 0.3)
    ax.axvline(12, color=vs.MUTED, lw=1, ls=":", zorder=1)
    ax.set_title(title, color=vs.INK)
    vs.style_axes(ax)
    ax.grid(axis="y", visible=False)
axes[-1].set_xlim(0, 22)
axes[-1].set_xlabel("CPP / CPPS (dB)")
handles = [
    plt.Line2D([], [], marker="D", ls="", color=vs.INK, ms=6, label="ROC cutoff"),
    plt.Line2D([], [], marker="o", ls="", mfc="white", mec=vs.INK, mew=1.5, ms=6,
               label="Normative lower limit (2 SD below mean)"),
    plt.Line2D([], [], marker="s", ls="", color=vs.ORANGE, ms=6, label="ADSV"),
    plt.Line2D([], [], marker="s", ls="", color=vs.BLUE, ms=6, label="Praat"),
]
fig.legend(handles=handles, loc="lower center", ncol=4, frameon=False, fontsize=8,
           bbox_to_anchor=(0.55, 0.0))
fig.tight_layout(rect=(0, 0.05, 1, 1))
plt.show()

Figure 4.1: Published CPP reference values for the Rainbow Passage (left) and sustained /a/ (right), by program. Filled diamonds are ROC cutoffs between voice-disordered and typical speakers; open circles are normative lower limits for typical speakers (2 SD below the mean). Each label names the study and, where it matters, the settings or the portion of the passage analysed. The dotted line at 12 dB is illustrative: on sustained /a/, a value of 12 dB is above the ADSV cutoff (11.46 dB) and below the Praat cutoff (14.45 dB), so the same number points in opposite directions. ADSV versions are not stated by Sauder et al. or Buckley et al., so those points illustrate the program difference and are not reference values. Sources: Murton et al. (2020); Sauder et al. (2017); Buckley et al. (2023).

4.6 Evidence 4: composite indices across languages

Composite indices add language to the pipeline, because their continuous-speech part is language-specific. The AVQI threshold differs between versions and languages (Barsties v. Latoszek et al. 2020), and the syllable count of the continuous-speech part is set separately for each language (Jayakumar and Benoy 2024). Figure 4.2 shows the published AVQI 03.01 thresholds. A meta-analysis of eight AVQI v03 studies found thresholds from 1.33 to 3.15 and did not derive a weighted threshold (Jayakumar and Benoy 2024); the US English validation, published after its July 2021 search, reports 1.17 (Castillo-Allendes et al. 2023). ABI thresholds in the ABI meta-analysis range from 2.94 (Brazilian Portuguese) to 3.69 (Korean) (Barsties v. Latoszek et al. 2021), and the US English validation reports 2.35 (Castillo-Allendes et al. 2023). Within one language the task matters too: in the same 53 Brazilian Portuguese speakers, the ABI threshold was 2.38 for counting and 3.13 for reading text, and the AVQI threshold 1.16 and 1.56 (Englert et al. 2020).

avqi = [  # language, threshold
    ("US English", 1.17),
    ("Brazilian Portuguese", 1.33),
    ("German", 1.85),
    ("Japanese (second-hand)", 2.06),
    ("Spanish", 2.28),
    ("French", 2.33),
    ("Dutch", 2.43),
]
fig, ax = plt.subplots(figsize=(7, 2.9))
ax.axvspan(1.33, 3.15, color=vs.GRID, alpha=0.8, lw=0)
ax.text(3.13, len(avqi) - 0.45, "meta-analytic range, 8 studies to 2021", ha="right",
        va="center", fontsize=8, color=vs.INK)
y = np.arange(len(avqi))
for yi, (lang, thr) in zip(y, avqi):
    if "second-hand" in lang:
        ax.plot(thr, yi, "o", ms=7, mfc="white", mec=vs.BLUE, mew=1.8, zorder=3)
    else:
        ax.plot(thr, yi, "o", ms=7, color=vs.BLUE, zorder=3)
    ax.text(thr + 0.04, yi, f"{thr:.2f}", va="center", fontsize=8, color=vs.INK)
ax.set_yticks(y)
ax.set_yticklabels([l for l, _ in avqi])
ax.set_xlim(1.0, 3.3)
ax.set_ylim(-0.6, len(avqi) - 0.1)
ax.set_xlabel("AVQI 03.01 threshold")
vs.style_axes(ax)
ax.grid(axis="y", visible=False)
fig.tight_layout()
plt.show()

Figure 4.2: Published AVQI 03.01 thresholds by language (points) and the range reported across eight AVQI v03 studies in a meta-analysis (shaded band, 1.33–3.15). The meta-analysis did not derive a single weighted threshold. The US English study was published after the meta-analysis search. The open point (Japanese) is reported second-hand, without a stated pipeline, and is not used as a cutoff. Sources: Castillo-Allendes et al. (2023); Englert et al. (2021); Barsties v. Latoszek et al. (2020); Delgado Hernández et al. (2018); Pommée et al. (2020); Barsties & Maryn (2016); Jayakumar & Benoy (2024).

The sources disagree on why the thresholds differ (see Points of disagreement), but not on the practical consequence: a threshold validated in one language is not a threshold for another.

4.7 Evidence 5: one script, different program versions

Even a fixed script is not a fixed pipeline, because the program underneath it changes. Stappenbeck et al. (2020) ran the AVQI 03.01 and ABI scripts on the same 218 German recordings under seven Praat versions. Five were versions cited in validation studies; the other two were 6.0.46 and 6.0.48, the most recent at the start of the study. The versions formed three clusters: 5.3.55, 5.3.57 and 6.0.06; 6.0.21, 6.0.22 and 6.0.48; and 6.0.46 alone, which carries a bug in the CPPS computation introduced in 6.0.44 and removed in 6.0.47 (Stappenbeck et al. 2020).

clusters = ["5.3.55 / 5.3.57 / 6.0.06", "6.0.21 / 6.0.22 / 6.0.48", "6.0.46 (CPPS bug)"]
sens = {"AVQI": [72, 71, 23], "ABI": [72, 70, 9]}
fig, ax = plt.subplots(figsize=(7, 2.4))
y = np.arange(len(clusters))[::-1]
for (name, vals), col, dy in zip(sens.items(), (vs.BLUE, vs.ORANGE), (0.12, -0.12)):
    ax.plot(vals, y + dy, "o", ms=7, color=col, label=name, zorder=3)
    for v, yi in zip(vals, y + dy):
        ax.text(v + 2, yi, f"{v}%", va="center", fontsize=8, color=vs.INK)
ax.set_yticks(y)
ax.set_yticklabels(clusters)
ax.set_xlim(0, 100)
ax.set_ylim(-0.6, len(clusters) - 0.4)
ax.set_xlabel("Sensitivity at the German threshold (%)")
ax.legend(frameon=False, fontsize=8, loc="lower right")
vs.style_axes(ax)
ax.grid(axis="y", visible=False)
fig.tight_layout()
plt.show()

Figure 4.3: Sensitivity of AVQI and ABI at the German thresholds (AVQI 1.85, ABI 3.42) when the same scripts run on the same 218 recordings under three clusters of Praat versions. The cluster containing 6.0.46, with its CPPS bug, misses most dysphonic voices, while specificity rises to 100% and correlations with perceptual ratings stay at 0.84–0.86. Source: Stappenbeck et al. (2020).

The bug lowered AVQI scores by about 4.2 points on average, so most dysphonic voices fell below the threshold (Figure 4.3). The correlations with perceptual ratings and the AUC stayed high in all three clusters, so validity statistics alone would not have revealed the problem (Stappenbeck et al. 2020). Between the two unaffected clusters, mean differences were not significant, but single recordings differed by up to about 0.6 AVQI points and 3 ABI points. The authors suspect a further change between 6.0.06 and 6.0.21, and recommend either the Praat version of the validation study or a check of each new version against an earlier one on test recordings (Stappenbeck et al. 2020).

4.8 How this reference reports reference values

Every Normative data table in the measure chapters carries two provenance columns.

• Pipeline: the program, version and non-default settings as stated by the source (“not stated” when the source does not say), the speech material, and the language where the measure depends on it.
• Class: how the value may be used, in one of four compatibility classes, or unclassified.

A measured value may be compared with a reference only when the pipelines are compatible. For physical quantities, such as F0 in Hz or a duration, that means the same quantity, the same task and a documented method. For algorithm-defined measures, such as jitter, shimmer, HNR, CPP, AVQI or ABI, it means the same algorithm and settings, the same software version and the same task, and for language-dependent indices the same language.

Table 4.2: Compatibility classes used in Normative data tables.

Class	Meaning	Use
pipeline-identical	Computed with the analysing software’s own engine on a reference database	Direct comparison. No value in this release is in this class yet.
pipeline-compatible	A published value whose stated pipeline meets the compatibility rule	Compare only measurements made with the same pipeline
language-conditional	A cutoff validated in a named language	Use for that language, task and stated pipeline; no cutoff is endorsed for a language without its own validation
meta-analytic	Pooled across studies or analysis systems	Descriptive context only, with its heterogeneity; never a cutoff, including pooled “weighted” thresholds
unclassified	The source does not state enough of the pipeline to place the value	Shown only as history, labelled as such; never a cutoff

Where the reviewed sources give no usable reference for a population, task or language, the table says so instead of borrowing a value from another pipeline. The measured value can still be reported; what is withheld is the comparison.

4.9 Using a reference value

1. Match the pipeline. Compare CPP data only with data obtained using the identical program and algorithm (Watts et al. 2017); compare ADSV values with ADSV norms and Praat values with Praat norms (Buckley et al. 2023).
2. Match the material and the language. Check the task, the portion of the passage, and, for composite indices, the language of the validation.
3. Report the pipeline with every value. The 1994 workshop consensus asks that the perturbation function and measure be made explicit in any perturbation analysis, and that F0, intensity and voice quality be defined when differences are reported (Titze 1995). The Praat version should be reported for CPPS (Buckley et al. 2023).
4. Do not convert across programs without a published conversion for your exact pipeline. Watts and colleagues give regression equations from ADSV CPP to Praat CPPS, separately for each language and task (Watts et al. 2017); they apply only to those settings and stimuli.

4.10 Caveats

• High correlation does not mean interchangeable values. Even correlations close to 1.0 can leave nontrivial residual error (Watts et al. 2017).
• Opaque software. Buckley and colleagues note that ADSV’s voicing detection is not detailed in its user manual (Buckley et al. 2023).
• Recording conditions. Hardware, microphone placement, environmental noise and software are known to affect perturbation measures; their effect on CPP and CPPS remains unclear (Barsties v. Latoszek et al. 2018).
• Explanations are hypotheses. Buckley and colleagues suggest their wider age range may explain their lower values (Buckley et al. 2023), although age group had no significant effect in their own data.
• Inter-laboratory standards. Until they are agreed, comparison of data sets must be undertaken with caution (Baken and Orlikoff 2000, 194).

4.11 Points of disagreement

• Why AVQI thresholds differ between languages. Englert et al. (2021) suggest that a low threshold may be a characteristic of Brazilian Portuguese, and that Brazilian listeners may be less tolerant of vocal deviation. Barsties v. Latoszek et al. (2020) relate the low German threshold to the small syllable count of its continuous-speech part. Jayakumar and Benoy (2024) suggest extreme thresholds could be due to the standardized syllable number; their own tables list the highest v03 threshold, 3.15, for a Korean study.
• Sauder’s 19.10 dB. Buckley et al. (2023) describe 19.10 dB as the average CPPS in Sauder et al. (2017). In the original it is the ROC cutoff; the mean for speakers without voice disorders was 20.11 dB.

Baken, Ronald J., and Robert F. Orlikoff. 2000. Clinical Measurement of Speech and Voice. 2nd ed. Singular Thomson Learning.

Barsties v. Latoszek, Ben, Geun-Hyo Kim, Jonathan Delgado Hernández, et al. 2021. “The Validity of the Acoustic Breathiness Index in the Evaluation of Breathy Voice Quality: A Meta-Analysis.” Clinical Otolaryngology 46 (1): 31–40. https://doi.org/10.1111/coa.13629.

Barsties v. Latoszek, Ben, Bernhard Lehnert, and Ben Janotte. 2020. “Validation of the Acoustic Voice Quality Index Version 03.01 and Acoustic Breathiness Index in German.” Journal of Voice 34 (1): 157.e17–25. https://doi.org/10.1016/j.jvoice.2018.07.026.

Barsties v. Latoszek, Ben, Youri Maryn, Ellen Gerrits, and Marc De Bodt. 2018. “A Meta-Analysis: Acoustic Measurement of Roughness and Breathiness.” Journal of Speech, Language, and Hearing Research 61 (2): 298–323. https://doi.org/10.1044/2017_JSLHR-S-16-0188.

Bielamowicz, Steven, Jody Kreiman, Bruce R. Gerratt, Marc S. Dauer, and Gerald S. Berke. 1996. “Comparison of Voice Analysis Systems for Perturbation Measurement.” Journal of Speech and Hearing Research 39 (1): 126–34. https://doi.org/10.1044/jshr.3901.126.

Buckley, Daniel P., Defne Abur, and Cara E. Stepp. 2023. “Normative Values of Cepstral Peak Prominence Measures in Typical Speakers by Sex, Speech Stimuli, and Software Type Across the Life Span.” American Journal of Speech-Language Pathology 32 (4): 1565–77. https://doi.org/10.1044/2023_AJSLP-22-00264.

Castillo-Allendes, Adrián, Juliana Codino, Lady Catherine Cantor-Cutiva, et al. 2023. “Clinical Utility and Validation of the Acoustic Voice Quality and Acoustic Breathiness Indexes for Voice Disorder Assessment in English Speakers.” Journal of Clinical Medicine 12 (24): 7679. https://doi.org/10.3390/jcm12247679.

Englert, Marina, Ben Barsties v. Latoszek, Youri Maryn, and Mara Behlau. 2021. “Validation of the Acoustic Voice Quality Index, Version 03.01, to the Brazilian Portuguese Language.” Journal of Voice 35 (1): 160.e15–21. https://doi.org/10.1016/j.jvoice.2019.07.024.

Englert, Marina, Livia Lima, and Mara Behlau. 2020. “Acoustic Voice Quality Index and Acoustic Breathiness Index: Analysis with Different Speech Material in the Brazilian Portuguese.” Journal of Voice 34 (5): 810.e11–17. https://doi.org/10.1016/j.jvoice.2019.03.015.

Jayakumar, Thirunavukkarasu, and Jesnu Jose Benoy. 2024. “Acoustic Voice Quality Index (AVQI) in the Measurement of Voice Quality: A Systematic Review and Meta-Analysis.” Journal of Voice 38 (5): 1055–69. https://doi.org/10.1016/j.jvoice.2022.03.018.

Karnell, Michael P., Kelly Dailey Hall, and Karen L. Landahl. 1995. “Comparison of Fundamental Frequency and Perturbation Measurements Among Three Analysis Systems.” Journal of Voice 9 (4): 383–93. https://doi.org/10.1016/S0892-1997(05)80200-0.

Kent, Raymond D., and Charles Read. 2002. The Acoustic Analysis of Speech. 2nd ed. Singular Thomson Learning.

Maryn, Youri, Paul Corthals, Paul Van Cauwenberge, Nelson Roy, and Marc De Bodt. 2010. “Toward Improved Ecological Validity in the Acoustic Measurement of Overall Voice Quality: Combining Continuous Speech and Sustained Vowels.” Journal of Voice 24 (5): 540–55. https://doi.org/10.1016/j.jvoice.2008.12.014.

Murton, Olivia, Robert Hillman, and Daryush Mehta. 2020. “Cepstral Peak Prominence Values for Clinical Voice Evaluation.” American Journal of Speech-Language Pathology 29 (3): 1596–607. https://doi.org/10.1044/2020_AJSLP-20-00001.

Pommée, Timothy, Youri Maryn, Camille Finck, and Dominique Morsomme. 2020. “Validation of the Acoustic Voice Quality Index, Version 03.01, in French.” Journal of Voice 34 (4): 646.e11–26. https://doi.org/10.1016/j.jvoice.2018.12.008.

Sauder, Cara, Michelle Bretl, and Tanya Eadie. 2017. “Predicting Voice Disorder Status from Smoothed Measures of Cepstral Peak Prominence Using Praat and Analysis of Dysphonia in Speech and Voice (ADSV).” Journal of Voice 31 (5): 557–66. https://doi.org/10.1016/j.jvoice.2017.01.006.

Stappenbeck, Lydia, Ben Barsties v. Latoszek, Ben Janotte, and Bernhard Lehnert. 2020. “Acoustic Voice Quality Index and Acoustic Breathiness Index as Two Examples for Strengths and Weaknesses of Free Software in Medicine.” Biomedical Signal Processing and Control 59: 101938. https://doi.org/10.1016/j.bspc.2020.101938.

Titze, Ingo R. 1995. Workshop on Acoustic Voice Analysis: Summary Statement. National Center for Voice; Speech.

Watts, Christopher R., Shaheen N. Awan, and Youri Maryn. 2017. “A Comparison of Cepstral Peak Prominence Measures from Two Acoustic Analysis Programs.” Journal of Voice 31 (3): 387.e1–10. https://doi.org/10.1016/j.jvoice.2016.09.012.