Acoustic Voice Assessment via Telehealth: What Survives the Compression

Videoconferencing platforms are optimized for intelligible conversation, not for high-fidelity acoustic capture. To deliver clear speech under variable bandwidth, every major platform applies a chain of processing before the audio reaches the other end:

• Lossy, bandwidth-limited coding. Voice-over-IP codecs discard information judged perceptually unimportant and reduce spectral bandwidth when the connection is constrained. Weerathunge et al. (2021) found that both uplink and downlink speeds predicted the size of measurement errors for spectral and cepstral measures.
• Noise suppression. Algorithms attenuate what they classify as non-speech. Sustained vowels look, to a suppressor, a lot like steady-state noise: Weerathunge et al. observed cases where a platform suppressed sustained-vowel amplitude almost entirely, and other cases where the envelope was flattened to a constant peak.
• Automatic gain control and level normalization. Level-dependent gain rewrites the amplitude envelope. This does not simply “compress dynamics”: in the 2021 data, Cisco WebEx and Zoom-with-enhancements inflated the SPL range of connected speech, while Zoom with enhancements turned off reduced it because of added noise.
• Echo cancellation and de-reverberation. Nonlinear processing that changes spectral content, increasingly performed jointly with noise suppression by a single neural network.

Each of these processes is helpful for conversation and harmful for measurement. The result is a signal that sounds like the patient's voice but is acoustically a different object than what the microphone captured.

The most systematic study of this question remains Weerathunge, Segina, Tracy, and Stepp (2021). They took recordings from 29 speakers with dysphonia (Parkinson's disease, muscle tension dysphonia, adductor laryngeal dystonia, nodules, polyp, scar, paralysis; CAPE-V overall severity 0–64 mm), originally captured in a sound booth with a headset microphone at 44.1 kHz, and transmitted each one through six HIPAA-compliant configurations: Zoom with enhancements, Zoom without enhancements (“original sound”), Cisco WebEx, Microsoft Teams, Doxy.me, and VSee Messenger. The received audio was recorded at the far end and re-analyzed in Praat.

Two design details matter for interpretation. First, samples were played through a loudspeaker 58 cm from a laptop's built-in microphone, in three real home environments, with the computers' own audio enhancements switched off. That isolates the platform's effect but also represents a patient sitting at a laptop without a headset, which is a common but not optimal setup. Second, the authors set significance at p < .005 to control for multiple comparisons, so “not significant” is a conservative statement.

The measures were mean f0, f0 standard deviation, f0 range, SPL range (connected speech), HNR (sustained vowel), L/H spectral ratio and CPPS (both sustained vowel and connected speech), and relative fundamental frequency. Here is what happened to each:

How big is big? Weerathunge et al. put the shifts next to published between-group differences. The 5.4–8.8 dB drop in HNR dwarfs the roughly 1 dB difference reported between typical and dysphonic voices; the same is true for L/H ratio. CPPS in connected speech moved by 1.4–2.3 dB, which is less than the 2.62 dB separation between speakers with and without voice disorders reported by Sauder, Bretl, and Eadie (2017). The authors' own conclusion is that f0 measures and CPPS are the acoustic measures most likely to carry clinically relevant information over telepractice. Note the direction: every platform pushed CPPS down, toward the dysphonic side of any cutoff. A live-capture CPPS compared against a normative threshold will over-call dysphonia.

Platforms and settings. Microsoft Teams had the fewest effects: f0 measures and SPL range were not significantly altered, although CPPS and HNR still fell with large effect sizes. Zoom with enhancements had the most pronounced effects overall. Turning Zoom's enhancements off (“original sound”) did not rescue the noise-based measures: CPPS and HNR still dropped with large effects, and the un-suppressed signal carried enough ambient noise to reduce the SPL range. WebEx, VSee, and Doxy.me artificially sustained the amplitude of vowels, so connected-speech measures were less affected than sustained-vowel measures on those platforms.

Predictors of error. Ambient noise at the transmitting end was a significant predictor of the difference between transmitted and original values for every measure except f0. Internet speed mattered too: uplink speed predicted errors in L/H ratio and vowel CPPS, and downlink speed predicted errors in f0 mean and SD, SPL range, and CPPS. Voice severity also predicted error for CPPS, HNR, and f0 range, meaning the patients whose voices you most need to measure are the ones whose signals are most degraded. An earlier version of this guide stated that the room mattered more than the connection; the data support “both matter,” with the room the more controllable of the two.

The fundamental issue with live videoconferencing is that the audio is processed in real time by the platform before it reaches the clinician. The signal that arrives is not the signal the patient produced.

The workaround is to bypass the platform entirely for the recording itself. The patient records the voice sample locally on their own device, using an app that saves an uncompressed or losslessly compressed file (WAV, FLAC, or Apple Lossless), then uploads the file separately through a secure portal or compliant file transfer. The clinician analyzes the uploaded file rather than the live transmission. Weerathunge et al. recommended exactly this in their clinical implications.

The evidence that local recordings on consumer devices are good enough has firmed up considerably:

• Bridge2AI-Voice validation (Awan et al., 2024). Across a diverse corpus of typical and disordered voices, CPP obtained from smartphones and tablets, with and without a low-cost headset microphone, correlated with a research-standard chain (GRAS 40AF microphone at 2.5 cm in a sound-treated booth) at a mean r of .99 (range .98–.99). Recording method still produced statistically significant offsets, and the L/H ratio was strongly affected by microphone type, but the relationships were linear and therefore correctable. Correlation is not identity: a phone CPPS is not a booth CPPS, and normative comparisons still require device-aware, pipeline-matched references.
• Smartphones versus the gold standard (Awan et al., 2025). iPhone 12 and SE and Samsung S21 and S9, at 15 and 30 cm, in a booth and in an ordinary quiet office, produced cepstral and spectral measures comparable to a flat-response reference microphone across 24 speakers spanning sex, age, f0, and voice-quality types.
• Meta-analysis (Barsties v. Latoszek et al., 2025). Ten studies, 379 participants, Apple and Samsung devices against clinical recording systems for jitter, shimmer, HNR, CPPS, and AVQI. The pooled picture is favorable but not uniform: the authors caution that for some parameters, current smartphone recordings do not yet match the precision of a clinical recording system. Treat perturbation measures from phones with more suspicion than cepstral ones.
• Clinical implementation (Schneider et al., 2024). The UCSF Voice and Swallowing Center attempted remote recordings on 108 patients over six months with a phone app, and documented the process questions that matter in practice: which normative data to compare against, how to train clinicians, and which patient-side problems actually occur.
• Formants (Zhang et al., 2021). In simultaneous recordings, f0 was tracked accurately by a field recorder, by lossless phone apps, and by Zoom, but formant estimates from phone apps were much closer to the reference than those from Zoom. If you use vowel-space or formant measures, they belong in the uploaded file, not the live session.

Asynchronous capture also has practical advantages that are independent of the fidelity argument: the patient can wear a headset at a fixed distance, the clinician can listen to the file before computing anything and request a re-recording if there is clipping or background noise, and each session uses the same device and protocol, which removes platform variability from longitudinal comparisons.

For a clinician integrating acoustic assessment into a telepractice voice caseload, the following workflow balances measurement quality against patient burden:

Asynchronous recording is the right default, but live capture has legitimate uses when the limitations are understood:

• Mean f0 for biofeedback, including gender-affirming voice work. Mean f0 was preserved across platforms and populations; within-session pitch tracking during practice is valuable and does not require absolute accuracy. Do not extend this to f0 variability, which was inflated on most platforms.
• Auditory-perceptual evaluation. Expert CAPE-V ratings on transmitted audio were not substantially degraded (Dahl et al., 2021). Rate live; measure from the file.
• Initial triage when no asynchronous option exists. A live impression is better than none, provided the chart records that the assessment was platform-mediated and that acoustic indices were not collected.
• Relative change within one session. Showing a patient their pitch or loudness trace in real time during a session is informative as a relative display, even though the absolute values are not.

The line to draw is between using the platform's audio for clinical interaction and quantifying the platform's audio as if it were a clean recording. The first is fine; the second is not.

Telepractice acoustic data needs to be defensible if it ends up in an outcome report, a peer-reviewed publication, or a third-party payer audit. Three documentation practices help:

• Describe the recording chain explicitly. “Patient recorded sustained /a/ and CAPE-V sentences on an iPhone with a wired boom headset at approximately 5 cm, Voice Memos at lossless quality with enhancement off, in a quiet home office, uploaded via [secure portal].”
• Use the same chain for baseline and follow-up. If the device or microphone changes, document it and treat the first recording on the new chain as a new reference point.
• Note any compromises. If a value was taken from live videoconferencing, say so and say what was not collected: “Mean f0 collected via Teams (version noted); CPPS, HNR, and AVQI not collected this session due to platform limitations.”

These practices make the methodology auditable and protect the clinician from later questions about whether an observed change reflects the voice or the measurement.