What VoxAtlas does

VoxAtlas is a quick, ~90-second voice check that looks for changes in your voice vs. your own baseline. It is not a diagnosis. When there’s a noticeable change, it indicates where the shift seems to come from:

• Laryngeal (voice source/quality: hoarseness, breathiness)
• Pulmonary (breath support/phrase length: pauses, phrasing)
• Neuro (timing/coordination: “pa-ta-ka” regularity)
• Affect (prosody: pitch variability / monotony)

What’s novel here

• Within-person, circadian-aware baselining. We compare you to you, and we bucket history by time-of-day to respect normal daily drift.
• Compact, multi-task protocol. Quiet → sustained “aaah” → DDK (“pa-ta-ka”) → short read → cough/sniff → one sentence—broad coverage in < 2 minutes.
• Transparent feedback. You see audio quality, system scores (0–1), and feature z-scores once a baseline exists.
• Privacy-first. Features are computed locally on your device/Pi; raw audio isn’t saved unless you opt in.

Signals we analyze (at a glance)

• Source/quality: spectral flatness, spectral centroid/roll-off; F0 (pitch) mean & variability (pYIN).
• Timing/breathing: pause ratio, articulation/onset rate, breath-group cues from the short read.
• Neuromotor: DDK rate and irregularity from “pa-ta-ka”.

We then compute robust z-scores vs your baseline (by hour-bin) and summarize which system changed.

How your data are used

• On device: features are computed in the browser and on the Pi.
• By default, no raw audio is stored. Only numeric features + session summaries are kept locally to build your baseline.
• Opt-in audio saving: enable if you want better troubleshooting/research; you can delete the local database anytime.
• Network: If you enable external access (e.g., Tailscale Funnel), traffic is HTTPS/TLS. No analytics, no third-party trackers.

Example result (with baseline)

Status: Watch (confidence 0.82). Small change from your usual morning voice, mainly in the laryngeal signal. Quality OK (SNR 23 dB, clipping 0.2%).

• Why: ↑ spectral flatness (+1.6σ), slight ↑ DDK irregularity (+0.6σ).
• What to do: Re-check tomorrow. If you feel unwell or it persists, consider a medical check.

VoxAtlas is a research prototype and not a medical device.

Selected references

• Mauch & Dixon. pYIN: A Fundamental Frequency Estimator Using Probabilistic Threshold Distributions. ICASSP 2014. PDF
• Bhattacharya et al. Coswara: A respiratory sounds and symptoms dataset. Scientific Data, 2023. Article
• Allison et al. Automated diadochokinesis analysis (DDK). 2022. PMC
• Tanchip et al. Validating automatic DDK methods in ALS. 2022. PMC
• Garrett & Healey. Fluctuations in voices of normal speakers across the day. 1987. PubMed
• Zacharia et al. Effect of circadian cycle on voice. 2018. Journal
• Dubnov. Generalization of Spectral Flatness Measure. (tech report). PDF
• Zeng et al. Acoustic & prosodic features in lung-cancer speech (pause ratio, etc.). 2023. PMC