Acoustic Voice Assessment End to End: The ASHA-Recommended Protocol in Practice

October 25, 202624 min readJorge C. Lucero

🎯 Key Takeaways

The core acoustic set is smaller than most parameter lists. Habitual, maximum and minimum SPL; mean f0, f0 SD, maximum and minimum f0; and CPP reported separately for sustained vowel and connected speech. That is the protocol's recommended set.
Jitter and shimmer are not in it. The panel describes them as valid only for mild-to-moderate dysphonia on long, steady vowels, and admits them only if the user recognises those limits. CPP was chosen as the general measure instead.
Recording setup is the irreducible part. Microphone type, distance, gain, sampling rate and room noise bound what any analysis can recover. No software removes this step, and browser-based analysis does not change the physics.
Absolute SPL requires calibration software cannot infer. A recording stores level relative to its own digital scale. Converting to dB SPL needs a sound level meter at a known distance — 30 cm, or 10 cm with a 9.5 dB correction.
The protocol contains no normative values. It standardises how to measure, not what counts as normal. Every typical-range boundary you see came from somewhere else, and the source has to be named.
Values are scale-bound. The panel states that instrumental measures cannot be compared across instruments and algorithms. CPPS, CSID and the DSI each illustrate this in a different way.

Instrumental acoustic assessment is a recommended component of voice evaluation, and has been since the American Speech-Language-Hearing Association convened an expert panel to specify how it should be done. That panel's report — Patel et al. (2018) — is the closest thing the field has to a protocol standard. It specifies the recording chain, the voice tasks, the measures to report, and the conditions that must be documented alongside them.

What it does not do is tell you how to carry it out in any particular software. That gap is where clinical acoustic analysis usually goes wrong: not in the measurement theory, but in the dozen configuration decisions between a patient in a chair and a number in a chart.

This guide walks the protocol end to end, organised the way an assessment runs — setup, calibration, recording, analysis, documentation. At each step it states what the protocol requires, how to satisfy it in PhonaLab, and where PhonaLab reports nothing because nothing can be justified. The last of those is the part worth reading closely.

What “ASHA-recommended” means here

The protocol is the recommendation. The ASHA expert panel recommends a set of tasks, measures and recording conditions; it does not endorse, certify or evaluate any software, including this one. Where this guide says a step follows the protocol, it means the step implements what the panel specified — a claim about the method, checkable against the paper cited at the end.

Part 1: What the Protocol Actually Asks For

Start here, because the core measure set is narrower than the parameter lists most acoustic software presents, and the omissions are deliberate.

The panel recommends reporting:

Vocal sound level — habitual SPL in dB, plus maximum and minimum SPL. The maximum SPL range may also be derived.
Vocal frequency — mean f0, f0 standard deviation, and maximum and minimum f0 in Hz. The phonatory f0 range may additionally be expressed in semitones.
Noise and quality — cepstral peak prominence (Hillenbrand & Houde, 1996), reported separately for the sustained vowel (CPP_vowel) and for connected speech (CPP_speech).

Jitter and shimmer are not in the core set

The panel selected CPP as its general measure of dysphonia because it reflects the overall balance of periodic and aperiodic energy, and because cepstral measures remain informative across the full range of severity and in connected speech as well as sustained vowels. Jitter and shimmer are described as valid only for mild-to-moderate dysphonia and only on long, steady sustained vowels. They may be added if the user recognises those limitations — but they are an addition, not a requirement, and a report that leads with them is not leading with what the protocol prioritises.

This matters for how you read PhonaLab's Voice Analyzer output, which does report jitter and shimmer. They are there because clinicians and the existing literature use them, not because they carry the protocol's endorsement. For the reasoning in detail, see Jitter and Shimmer: What They Really Tell You.

Two further points about the scope of the document. First, the protocol contains no normative values. It standardises measurement and leaves the question of what a value means to the separate literature on reference data. Second, the panel states plainly that instrumental measures cannot be compared across different instruments and algorithms. That single sentence governs most of the caveats in Part 6.

Part 2: Recording Setup — The Part No Software Removes

Browser-based analysis eliminates installation and configuration. It does not eliminate physics. Everything in this section happens before any file reaches any analysis engine, and it bounds what the analysis can recover.

Microphone

A head-mounted omnidirectional microphone, positioned 4–10 cm from the lips at 45°–90° off the front of the mouth. Headset microphones are strongly recommended. The response should be flat within 2 dB from roughly 50 to 8000 Hz, with a noise floor at least 10 dB below the quietest sound to be recorded and enough headroom that the loudest phonation does not clip. Unidirectional microphones suppress room noise but can compromise SPL and spectral measures through the proximity effect.

Preamplifier

Specifications at least matching the microphone, with no equalisation, no tone controls, no automatic gain control and no noise cancellation. This is the most commonly violated requirement in practice, because consumer capture paths enable exactly these by default.

Digitisation

Minimum 44.1 kHz sampling rate and 16-bit resolution, with 24-bit preferred, through a high-quality external converter where available. Gain set so that loud phonations sit just below clipping. Uncompressed or lossless audio only — WAV rather than MP3 or AAC.

Environment

Ambient noise at least 10 dB below the quietest phonation, targeting below 38 dB C-weighted or unweighted, or below 25 dB A-weighted, measured at 30 cm. Signal-to-noise ratio of at least 30 dB, with more than 42 dB as the ideal. Minimise reverberation and intermittent noise.

All three processes the protocol bans act before digitisation, inside the capture path, which is why no later analysis can undo them: the file that reaches the analysis engine is already a different object from the sound at the microphone. Each one destroys a specific class of measure rather than degrading everything evenly — gain control flattens the level differences the loudness-range task exists to capture, and noise suppression attenuates exactly the aperiodic energy that cepstral and noise measures quantify, which is to say it removes the thing being measured. Consumer capture paths enable all three by default.

If the recording is not made in the room

Two situations break these requirements in ways worth understanding before interpreting anything. Audio transmitted live through a videoconferencing platform has already passed through coding, noise suppression and gain control; mean f0 largely survives this, while CPPS, HNR and intensity measures do not. Smartphone and tablet recordings, by contrast, can be clinically usable when made locally to an uncompressed file with a headset microphone at a fixed distance. The distinction is between using a platform's audio for clinical interaction and quantifying it as though it were a clean recording.

Detail in Acoustic Voice Assessment via Telehealth and Smartphone Voice Recording for Clinical Assessment.

Part 3: Calibrating for Sound Pressure Level

This is the step most acoustic workflows skip, and the reason the protocol's three intensity measures are the least reliably reported of its core set.

The problem is simple to state. A digital recording stores amplitude relative to its own full scale. Nothing in the file records how loud the sound actually was at the microphone, because that depends on microphone sensitivity, preamplifier gain, the converter and the distance from the mouth — none of which the file carries. Two recordings of the same phonation at different gain settings differ in digital level and not at all in dB SPL. Software that prints a dB figure straight from the waveform is reporting the former while implying the latter.

The conversion is a single additive offset for a fixed chain. It has to be measured, and measured again whenever anything in the chain changes.

The two sanctioned procedures

Preferred — meter at the reference distance. With the client sustaining /a:/, measure SPL with a sound level meter at 30 cm from the mouth, then adjust the recording system's reading to match. All reported SPL values are then referenced to 30 cm, and that distance is reported with them.
Alternative — meter beside a head-mounted microphone. Place the meter next to the headset microphone and convert using the distance law. The panel prefers a 10 cm mouth-to-microphone distance for this method, as it is more accurate than shorter distances, and gives a correction of 9.5 dB from 10 cm to 30 cm.

Use C-weighting on the meter. Results obtained under any other weighting are comparable only with data collected under that same weighting, which makes the weighting part of the measurement rather than a display setting. For the full treatment, including calibration of head-mounted microphones, see Švec and Granqvist (2018).

Distance discipline matters more than the decibel arithmetic

Sound pressure level falls by roughly 6 dB for each doubling of distance. A calibration made at 10 cm is invalidated by a microphone that drifts to 15 cm, and a headset re-seated between sessions is not a fixed mouth-to-microphone distance. This is why the protocol asks you to record the distance and gain for every recording and hold them constant across follow-up comparisons — the stability of the geometry, not the precision of the meter, is usually the limiting factor.

Note also that the close placements favoured for signal-to-noise in telehealth recording are not the placements favoured for intensity work. If you intend to report SPL, use the protocol's 10 cm with a documented correction, not a 2.5 cm headset position.

What PhonaLab does and does not report

PhonaLab accepts uploaded audio and cannot observe the recording chain that produced it. It therefore does not report habitual SPL, maximum SPL, minimum SPL or SPL range as absolute dB SPL values. Where level enters a computation, it does so as a relative quantity and is labelled as such.

The one place a calibration offset can currently be supplied is the Dysphonia Severity Index, whose I-Low term is weighted at −0.26 per decibel and is therefore sensitive to exactly this problem: an optional intensity calibration offset may be entered, and the level each recording was made at is reported beside the score. Absent an offset, nothing about the analysis changes and the score should be read as pipeline-relative.

This is a declared absence rather than a silent one. A borrowed cutoff or an uncalibrated number presented as dB SPL would be worse than no number, because it would travel into a chart looking like a measurement.

Part 4: The Recording Session

Four tasks, with specified durations and repetitions. Record each as its own uncompressed file, named so that the task is unambiguous — the analysis steps in Part 5 are per-task, and mismatching a file to a task is the easiest way to produce a confidently wrong result.

Task	What to elicit	Duration and repetitions
Sustained vowel	/a:/ at habitual pitch and loudness	3–5 s, three times
Standard reading passage	First paragraph of the Rainbow Passage (Fairbanks, 1960) for adults; “The Trip to the Zoo” for children who read	Whole passage where possible; otherwise a consistently chosen segment of at least 5 s
Loudness range	/a:/ as quietly as possible without whispering; /a:/ as loudly as possible	At least 2 s each, three times per task
Pitch range	/a:/ as high as possible including falsetto; /a:/ as low as possible in modal register without fry. Glides or stepwise tasks are also acceptable	At least 2 s each, three times per task

Why three repetitions — and why you must state which one you analysed

Maximum-performance tasks — loudest, quietest, highest, lowest — are partly learned within the session, so a single attempt risks understating the range and compressing every measure derived from it. The repetitions exist to let performance stabilise.

The protocol specifies the repetitions but does not specify which trial to carry forward into the analysis. That makes the selection rule yours, and therefore something to record: taking the best of three on a maximum-performance task and taking a representative trial on a habitual task are both defensible, and they are not comparable with each other. Fix one rule, write it in the chart, and apply it identically at follow-up.

Part 5: Running the Analysis

In a desktop analysis workflow this is the longest part: configure pitch floor and ceiling, set spectrogram and cepstrum parameters, confirm pulse detection settings, then either step through menu queries per measure or run a script that batches them. The configuration step is not optional, and getting it wrong changes the numbers without producing any error.

PhonaLab removes the configuration step by fixing it. Each measure is extracted by its own engine call at the parameterisation its reference publication used, with input guards that return could not calculate rather than a value when the input cannot support the measure. That is the substantive difference between the two workflows — not convenience, but the removal of a class of silent error.

Measures are computed in Python via Parselmouth, which interfaces the Praat signal-processing engine (Jadoul et al., 2018), and the platform's implementation has been validated against desktop Praat in a peer-reviewed study (Lucero et al., 2026).

Each recorded task reaches its protocol measures through the tool or tools that compute them, so the analysis is run per task rather than by uploading one file and reading everything off it. The reading passage feeds two: Spectral for CPP(speech), and the Pitch Visualizer for the f0 track, which is where f0 SD properly comes from. Three lanes complete; the loudness range task dead-ends, because absolute sound pressure level cannot be recovered from an uploaded file without the calibration described in Part 3. The AVQI/ABI path is drawn lighter because those indices sit outside the panel's core measure set, and it taps both the vowel and the passage because the script runs on a concatenation of the two.

Sustained vowel

Upload the habitual-pitch-and-loudness vowel to the Voice Analyzer. This returns the frequency and perturbation family — mean f0, f0 SD and f0 range via autocorrelation over a 60–500 Hz pitch range, jitter and shimmer from the voice-report pipeline, harmonics-to-noise ratio (Boersma, 1993), and glottal-to-noise excitation ratio (Michaelis et al., 1997) — together with CPPS in its Praat implementation (Maryn & Weenink, 2015), the protocol's CPP_vowel.

For the spectral family on the same file — long-term average spectrum slope and tilt, alpha ratio, H1–H2, high-frequency noise above 6 kHz, and CSID (Awan et al., 2016) — use the Spectral tool.

Connected speech

Upload the reading passage to the Spectral tool to obtain CPP_speech, and to the Pitch Visualizer for the f0 track, mean f0 and f0 standard deviation across the passage. Reporting CPP separately for vowel and speech is a protocol requirement rather than a refinement: the two values answer different questions and routinely diverge, which is much of why the panel asks for both.

Pitch range

Upload the highest and lowest productions, or the glide, to the Pitch Visualizer for maximum and minimum f0. Note that a pitch ceiling set below a speaker's actual maximum does not merely truncate the measurement — it can invert it, so that a higher voice scores lower on the measure built to capture the top of the range. This is why the glide analysis behind the DSI's F0-High term runs to a 1300 Hz ceiling rather than a conventional 500 or 800 Hz one.

Multiparametric indices

The AVQI/ABI tool computes the Acoustic Voice Quality Index (Maryn et al., 2010; Barsties & Maryn, 2016) and the Acoustic Breathiness Index (Barsties v. Latoszek et al., 2017) from a concatenated sustained-vowel and connected-speech sample, via the official Praat script with its published coefficients unmodified. AVQI requires a sampling rate of at least 22.05 kHz; where the input does not meet a requirement, the report states so explicitly rather than returning a degraded value.

Neither index is part of the panel's core set — the panel chose CPP as its single quality measure — but both are extensively validated against perceptual severity, and both run on recordings the protocol already asks you to collect. For what they add over single parameters, see AVQI and ABI Explained. For dimension-specific alternatives to a single severity score, see Why One Number Isn't Enough.

Part 6: Reading the Output Without Over-Reading It

The panel's warning that instrumental measures cannot be compared across instruments and algorithms is the most consequential sentence in the protocol, and the one most often ignored. Four specific instances are worth carrying into every report you read.

CPPS values live on a scale defined by their configuration

CPPS depends on both the analysis configuration and the software version. PhonaLab reports CPPS on the Praat/Maryn–Weenink scale, using the settings of the official AVQI/ABI script, and those values are comparable only with reference values obtained under the same configuration. Published Praat CPPS values in the approximate range 18–23 dB derive from a different configuration and are not comparable. A cutoff lifted from a paper without checking its configuration is a category error, not a rounding difference.

CSID is not interchangeable across platforms

PhonaLab's CSID is computed from Praat-derived CPPS and a spectral ratio. It is not interchangeable with CSID scores from ADSV/CSL and should not be read against severity bands published for those platforms. Concurrent validity against perceptual severity has been established for this implementation independently (Lucero et al., 2026), which makes the measure informative for tracking change within PhonaLab for the same speaker — and uninformative as a cross-platform number.

Three different numbers are called the DSI

Wuyts et al. (2000) fitted the index on CSL/MDVP instrumentation. Maryn, Morsomme and De Bodt (2017), porting it to Praat, refitted the weights rather than reusing the originals — their own script states that its output is not equal to the original. PhonaLab applies the original coefficients to its own measurements, which matches neither source. A DSI is interpretable against other scores produced the same way, and against nothing else. The full three-way comparison is in MPT and DSI: A Clinical Implementation Guide.

Input properties can disqualify a measure

Sampling rate is not a cosmetic property of a file: below a measure's threshold it is not computed at all, and the thresholds differ by measure (table below). PhonaLab reports which guards fired on each analysis, and long recordings are truncated to the configured analysis window with the truncation stated. Read those lines — they tell you what the numbers above them are made of. The GNE case is worked through in Glottal-to-Noise Excitation (GNE).

Measure	Minimum sampling rate	Below it
AVQI / ABI	22.05 kHz	Not computed; the report states that the requirement was not met.
GNE	10 kHz	Not computed; band-limited estimation is not reliable below this rate.
CPPS	Signal-dependent	Not computed where the analysis would fall above the signal's Nyquist frequency.
Protocol recommendation	44.1 kHz	Below the panel's recommended recording minimum. Measures above their own thresholds still compute, but the recording no longer meets the protocol.

On reference ranges

The protocol supplies none, so every typical-range indicator you see came from somewhere else and the somewhere matters. Two examples of what a sourced boundary looks like: CPPS below 13.33 dB is associated with clinically relevant dysphonia in a sustained vowel, derived on this platform's own pipeline (Lucero et al., 2026) and consistent with an independently reported cutoff of 14.45 dB (Murton et al., 2020); and the GNE typical-range floor of 0.87 is the tenth percentile of the healthy subset of the Perceptual Voice Qualities Database (Walden, 2022), computed on the same pipeline. Each PhonaLab report names the source of every reference value it applies, per measure, alongside the engine version. If a report shows you a boundary without a source, that is a defect in the report.

Part 7: Documenting and Comparing Across Sessions

The protocol's reporting requirements are not paperwork; they are what makes a second measurement comparable to a first.

Record the microphone distance and gain for every recording, and hold them constant for follow-up comparison such as pre- and post-treatment.
Report SPL with its reference distance and weighting — for example, 30 cm and C-weighting.
Report background noise level, using the same weighting as the SPL measures.
Include units, and use them consistently.
Use documented or validated algorithms, and state which, since values are not comparable across instruments.

PhonaLab's PDF report carries the last of these as a Methods and Validation section: the processing chain and engine version, each measure computed with its canonical method citation, the validation reference, the source of every reference value applied, and the transparency lines recording which guards fired or truncations were applied on that specific file. The first four remain yours to record, because they describe events in a room that no software observed.

For longitudinal work, Session Compare places analyses of the same speaker side by side — which is only meaningful if the recording conditions were held constant, which is why the documentation above comes first.

What This Protocol Does Not Cover Here

Stated plainly, so that nothing in this guide reads as a claim it does not make.

Element	Status
Habitual, maximum, minimum SPL and SPL range	Not reported as absolute dB SPL. Requires a calibrated chain PhonaLab cannot observe in an uploaded file. An optional intensity offset exists for the DSI only.
f0 range in semitones	Maximum and minimum f0 are reported in Hz; the semitone expression is an optional derivation in the protocol.
Laryngeal diadochokinesis, s/z ratio	Not implemented. Neither is part of the panel's core acoustic set, though both appear in wider assessment practice.
Aerodynamic measures — vital capacity, phonation quotient, mean flow rate	Out of scope. These are aerodynamic rather than acoustic and require instrumentation outside any acoustic analysis platform.

Summary

1Know the core set. SPL, f0, and CPP for vowel and speech separately. Jitter and shimmer are optional additions with stated limits, not the centre of the protocol.
2Fix the recording chain before anything else. Headset omnidirectional microphone at 4–10 cm, no gain control or noise processing, 44.1 kHz and 16-bit minimum, uncompressed, quiet room.
3Calibrate if you intend to report intensity, and hold the geometry. Meter at 30 cm, or at 10 cm with the 9.5 dB correction, C-weighted. Without this, intensity is relative and should be labelled so.
4Record four tasks with repetitions — habitual vowel, reading passage, loudness extremes, pitch extremes — as separate uncompressed files, and record which trial you analysed.
5Analyse per task, not per file dump. Voice Analyzer and Spectral for the vowel, Spectral and the Pitch Visualizer for the passage, the Pitch Visualizer for the range tasks, AVQI/ABI for the multiparametric indices.
6Treat every value as scale-bound. CPPS, CSID and the DSI are each interpretable only against values produced the same way. Check which guards fired.
7Document the conditions. Distance, gain, weighting, background noise, and algorithm. Without them a second measurement is a new measurement, not a comparison.

Run the protocol without a desktop install

PhonaLab analyses uploaded WAV recordings in the browser, with each measure extracted at its reference parameterisation and every reported value traced to a published source. Audio is processed in memory and never stored.

Open PhonaLab

References & Further Reading

Patel RR, Awan SN, Barkmeier-Kraemer J, Courey M, Deliyski D, Eadie T, Paul D, Švec JG, Hillman R. (2018). Recommended protocols for instrumental assessment of voice: American Speech-Language-Hearing Association expert panel to develop a protocol for instrumental assessment of vocal function. American Journal of Speech-Language Pathology, 27(3), 887–905. doi:10.1044/2018_AJSLP-17-0009
Švec JG, Granqvist S. (2018). Tutorial and guidelines on measurement of sound pressure level in voice and speech. Journal of Speech, Language, and Hearing Research, 61(3), 441–461. doi:10.1044/2017_JSLHR-S-17-0095
Fairbanks G. (1960). Voice and Articulation Drillbook (2nd ed.). Harper & Row. [Source of the Rainbow Passage]
Hillenbrand J, Houde RA. (1996). Acoustic correlates of breathy vocal quality: dysphonic voices and continuous speech. Journal of Speech and Hearing Research, 39(2), 311–321. doi:10.1044/jshr.3902.311
Maryn Y, Weenink D. (2015). Objective dysphonia measures in the program Praat: smoothed cepstral peak prominence and acoustic voice quality index. Journal of Voice, 29(1), 35–43.
Murton O, Hillman R, Mehta D. (2020). Cepstral peak prominence values for clinical voice evaluation. American Journal of Speech-Language Pathology, 29(3), 1596–1607.
Maryn Y, Corthals P, Van Cauwenberge P, Roy N, De Bodt M. (2010). Toward improved ecological validity in the acoustic measurement of overall voice quality. Journal of Voice, 24(5), 540–555.
Barsties B, Maryn Y. (2016). External validation of the Acoustic Voice Quality Index version 03.01 with extended representativity. Annals of Otology, Rhinology & Laryngology, 125(7), 571–583. doi:10.1177/0003489416636131
Barsties v. Latoszek B, Maryn Y, Gerrits E, De Bodt M. (2017). The Acoustic Breathiness Index. Journal of Voice, 31(4), 511.e11–511.e27.
Awan SN, Solomon NP, Helou LB, Stojadinovic A. (2016). Spectral-cepstral estimation of dysphonia severity: external validation. Annals of Otology, Rhinology & Laryngology, 125(2), 168–176.
Boersma P. (1993). Accurate short-term analysis of the fundamental frequency and the harmonics-to-noise ratio of a sampled sound. Proceedings of the Institute of Phonetic Sciences, 17, 97–110.
Michaelis D, Gramss T, Strube HW. (1997). Glottal-to-noise excitation ratio — a new measure for describing pathological voices. Acustica / Acta Acustica, 83, 700–706.
Wuyts FL, De Bodt MS, Molenberghs G, Remacle M, Heylen L, Millet B, Van Lierde K, Raes J, Van de Heyning PH. (2000). The Dysphonia Severity Index: an objective measure of vocal quality based on a multiparameter approach. Journal of Speech, Language, and Hearing Research, 43(3), 796–809. doi:10.1044/jslhr.4303.796
Maryn Y, Morsomme D, De Bodt M. (2017). Measuring the Dysphonia Severity Index (DSI) in the program Praat. Journal of Voice, 31(5), 644.e29–644.e40. doi:10.1016/j.jvoice.2017.01.002
Walden PR. (2022). Perceptual Voice Qualities Database (PVQD) (Version 4) [Dataset]. Mendeley Data. doi:10.17632/9dz247gnyb.4
Jadoul Y, Thompson B, de Boer B. (2018). Introducing Parselmouth: a Python interface to Praat. Journal of Phonetics, 71, 1–15. doi:10.1016/j.wocn.2018.07.001
Lucero JC, et al. (2026). Algorithm Verification and Concurrent Validity of a Web-Based Platform for Multiparametric Acoustic Voice Quality Indices. Journal of Voice (open access). doi:10.1016/j.jvoice.2026.04.009

⚠️ Educational Information

This article describes published measurement methods and the reference sources PhonaLab applies, for educational purposes. It does not constitute clinical advice and does not establish a diagnosis. The ASHA expert panel recommends a protocol; it does not endorse or certify any software. Clinical interpretation of acoustic measures remains the responsibility of the qualified, licensed clinician. PhonaLab provides acoustic measurement tools; it does not provide clinical interpretations or medical diagnoses.

Back to Guides