1. Purpose

Svara analyses a short, structured voice sample and reports quantitative speech markers that the literature associates with mood, motor-speech, and cognitive-linguistic function. It is intended to serve as a pre-consultation triage aid — a low-burden, repeatable prompt that may help prioritise which patients receive a full standardised assessment sooner. It supplements, and never replaces, clinical interview and validated instruments.

2. Assessment protocol

Total core protocol: approximately 90 seconds. Each task is chosen because different functions surface in different speech elicitations; a marker is only ever computed from a task that is valid for it.

Optional add-on tasks: diadochokinesis (pa-ta-ka) for motor-speech timing; semantic and phonemic verbal fluency; pitch glide; maximum phonation time; s/z ratio; story recall.

3. What is measured

Features are extracted with openSMILE eGeMAPSv02 (the INTERSPEECH standard parameter set), Praat voice-quality algorithms, and transcript-derived measures. Each marker is expressed as a z-score against a healthy reference distribution and reported as a percentile, so an "elevated" reading means N standard deviations from the reference rather than an arbitrary threshold.

4. Evidence base

Marker directions and relative weights are taken from the published literature rather than chosen by hand:

• Depression: reduced pitch variability is the most replicated acoustic marker (η² ≈ 0.18–0.31), with slower speech rate, longer and more frequent pauses, and reduced loudness dynamics.
• Anxiety: reduced word count and speaking duration, with shimmer and F0 associations. The evidence base here is substantially weaker — audio-only models in the literature reach only ~0.60 unweighted average recall, barely above the self-report reliability of GAD-7 itself. Anxiety output is labelled low-confidence accordingly.
• Motor speech / phonatory: elevated jitter and shimmer with reduced HNR in hypokinetic dysarthria; sustained phonation and DDK are the appropriate elicitations.
• Cognitive-linguistic: reduced lexical diversity, word-finding pauses, and reduced verbal fluency output; picture description is the standard elicitation.

5. Reference frame and reporting

Outputs are aligned to instruments already in routine use, as a reference frame:

6. Expected performance — read in context

6.1 Why published voice-AI figures mislead

Published voice-screening accuracies above 90% should be treated with scepticism. A 2026 controlled study demonstrated that such figures are largely produced by speaker leakage: when recordings from the same individual appear in both training and test sets, the model learns to recognise the speaker rather than the condition. In that study a model reporting 97.65% depression accuracy also achieved 90.95% speaker-identification accuracy; a comparable model without speaker information scored 62–67%. All figures below are strictly speaker-independent.

6.2 Measured performance of the phonatory model

Trained on a public sustained-phonation corpus (195 phonations, 32 speakers) and evaluated by leave-one-speaker-out cross-validation, with a 95% confidence interval bootstrapped by resampling speakers (rows within a speaker are correlated, so resampling rows would understate uncertainty):

6.3 What these numbers mean next to established instruments

An AUC in the 0.7–0.8 range reads as poor only if compared against diagnostic tests. Screening instruments in routine use operate in a similar range:

6.4 Operating point

The system is configured for high sensitivity at the cost of specificity (0.95 / 0.44 per clip), which is the appropriate posture for a triage aid: the objective is to avoid missing someone who warrants assessment, accepting that roughly half of flagged individuals will not have the condition. A positive result should therefore be expected to be a false positive much of the time, and carries no weight on its own — it indicates only that a standard assessment is worthwhile.

6.5 Expected performance in other domains

6.6 Self-report module (PHQ-8 and GAD-7)

The system administers PHQ-8 and GAD-7 after the voice tasks. These are validated, public-domain instruments and are reported alongside — not merged into — the voice indicators. Where the two disagree, the interface directs the patient according to the questionnaire, since it is the better-established measurement.

7. Limitations

• Not clinically validated. No patient-population validation has been performed.
• Reference norms are approximate until calibrated on a local healthy control group (≥30 participants recorded on the same equipment).
• Language dependence. Acoustic markers transfer across languages reasonably well; transcript-derived linguistic markers and the reading passage are English-specific. Results for other languages are not equivalent without recalibration.
• Confounds: sex and age (fundamental frequency is strongly sex-dependent), microphone and channel, upper respiratory infection, smoking, intoxication, sedating medication, and situational stress at the time of recording.
• Single time-point. A single recording is a snapshot; within-subject change over repeated sessions is more informative than any absolute value.
• State, not trait. Markers reflect current speech production, which varies day to day.

8. Suggested use in practice

1. Record in a quiet room; the system reports signal quality and rejects unusable audio.
2. Read the summary as a prompt, not a result.
3. Where an indicator is elevated, administer the aligned standardised instrument (PHQ-9 / GAD-7) and proceed with normal clinical assessment.
4. Where all indicators are low, this is not evidence of absence — a patient reporting distress should be assessed regardless of what the speech markers show.
5. Consider repeat recordings over time; trajectory is more meaningful than a single reading.

9. Data handling

Audio is processed on the server and used to compute the reported markers. All analysis runs locally on CPU; no audio is sent to any third-party service. Voice recordings constitute personal health information — the deploying institution is responsible for consent, retention, and applicable data protection obligations before any use with identifiable patients.

10. Selected references

Kroenke K et al. The PHQ-8 as a measure of current depression. J Affect Disord, 2009.
Spitzer RL et al. A brief measure for assessing generalized anxiety disorder: the GAD-7. Arch Intern Med, 2006.
Lovibond SH & Lovibond PF. Manual for the Depression Anxiety Stress Scales (DASS), 1995.
Cummins N et al. A review of depression and suicide risk assessment using speech analysis. Speech Communication, 2015.
The voice of depression: speech features as biomarkers for major depressive disorder. BMC Psychiatry, 2024.
Acoustic and linguistic features of impromptu speech and their association with anxiety. JMIR Mental Health, 2022.
Who is Speaking or Who is Depressed? A Controlled Study of Speaker Leakage in Speech-Based Depression Detection. arXiv, 2026.
Eyben F et al. The Geneva Minimalistic Acoustic Parameter Set (eGeMAPS). IEEE Trans Affective Computing, 2016.
Bridge2AI-Voice Consortium. An ethically-sourced, diverse voice dataset linked to health information. PhysioNet, 2025–26.