Hearing & tinnitus research

Development biography · earlier archive

Development profile compiled from project source and retained records in October 2026. Dates and test results describe those records; they do not imply a new build, live acceptance test or public release.

3 · Biography

Most consumer "hearing test" pages overstate themselves: they present a tone player as if it produced a clinical audiogram, ignore that browser audio is uncalibrated, and reduce tinnitus to a single matched pitch. The Hearing & Tinnitus Panel is the opposite wager — an instrument built around the honest question what can a browser actually measure, and how can it say when it doesn't know? Its documented intellectual centre (stated on the Methods page) is a specific, testable relationship: where does a person's tinnitus percept sit relative to their own measurable hearing loss — at the audiometric edge, inside the loss region, or at a displaced frequency — expressed in octaves and an estimated Greenwood tonotopic coordinate. That framing separates research-prototype capability (collecting repeatable within-person relationships) from clinical validation (calibrated dB HL, validated questionnaires), which the app never claims.

The intended audience is threefold: individuals who want to observe and track their own hearing/tinnitus between clinic visits; a producer/technical audience curious about auditory psychophysics; and an audiologist or ENT reading a shared session as context, not a result. The app is careful to be useful to each without misleading any.

What it does, concretely, spans four visible tabs plus a persistent "measurement spine". Quick gives a 1 kHz level-set, a 250 Hz→20 kHz hearing-range sweep, and a slow tinnitus pitch-match with a live oscilloscope + spectrum fine-tuner (with ÷2/×2 octave jumps to defeat octave confusion) averaged over up to three passes. Audiology runs per-ear pure-tone thresholds using warbled pulsed chirps and a Bayesian, ZEST-inspired adaptive estimator: each (ear, frequency) holds a posterior over its threshold, informed by priors from age and neighbouring frequencies/the other ear, and reports a threshold with a 95% credible interval and trial count rather than false single-dB precision. It adds catch trials (false-alarm rate), shuffled interleaving, a self-referenced downloadable audiogram with an illustrative ISO-7029-style age overlay, and a one-tap test–retest repeatability check. Speech-in-noise is a calibration-free digits-in-noise screen with a 1-up/1-down SNR track yielding a relative SRT. Research tests is a directly accessible exploratory suite: frequency-discrimination, gap-detection, temporal-modulation detection, binaural lateralisation, dichotic stream attention, a tinnitus spectrum/likeness map, a residual-inhibition function and a tinnitus tuning curve.

The distinctive ideas are less the individual tests than the discipline around them. A measurement spine ribbon runs above every tab: a protocol-wide safety state (sudden-hearing-change or neuro/vestibular symptoms switch the app to "seek care first"; sound-pain or uncertain tolerance switches to self-report-only and removes sound-provoking tasks for the session, reconfirmed on every load); a device/capability context block (AudioContext state, sample rate, stereo-routing check, a monitored master-bus RMS/peak/clipping in dBFS — explicitly device/signal facts, not acoustic calibration); momentary check-ins (prominence, loudness, distress, stress, tolerance) with local trends; an auditable requested/applied/observed event stream; and a persistent emergency-stop that silences every Web Audio voice and OS speech instantly. A Sound Lab offers bounded comfort sessions (narrowband/notched/pink/white maskers or a quiet-rest comparator) with fixed durations and before/after observations, and a prospective personal A/B comparison that computer-randomises balanced session pairs, conceals later allocations, and returns a descriptive within-person difference — never a p-value or efficacy claim.

A representative user journey: open the page on a phone with headphones; answer the safety check; set a comfort level; run the Quick sweep and pitch-match, or the full Audiology panel for a profile with uncertainty whiskers; optionally run a research test or a Sound Lab session; review everything on-screen. Results stay local in that browser by default. Nothing is transmitted on finishing a test — the user must explicitly export JSON/CSV, download the audiogram PNG, open a pre-filled email, or consent to an optional configured results service.

The honesty is structural, not a footnote: the profile is named a "self-referenced hearing profile", never a dB HL audiogram; speech-in-noise is an "SRT-equivalent (relative)"; the tinnitus "type" is a transparent rule-based suggestion that surfaces red flags (pulsatile, unilateral, sudden-onset-with-vertigo) and routes to an audiologist/ENT. Sound-therapy language is deliberately limited to exploratory research. The result is a tool that behaves like an instrument that knows the limits of its own measurements — which, for uncalibrated browser audio, is the whole point.

4 · Modules & features — what a visitor can see / hear / try

(All authored by the project in index.html; the only external moving part is the optional Netlify/Graph email path.)

• Quick test — reference-tone level set; hearing-range sweep; slow pitch-match with live oscilloscope + real-time spectrum + A/B-against-silence; rule-based tinnitus-character suggestion. See/hear/try.
• Audiology panel — per-ear warbled-chirp thresholds (250 Hz–8 kHz; optional 10–16 kHz EHF, flagged exploratory); downloadable PNG audiogram with CI whiskers + age overlay; device/stereo check; catch-trial false-alarm rate; test–retest repeatability; advanced tinnitus battery (pitch, sound-quality audition, loudness in dB SL, minimum masking level, MRIL residual-inhibition search, impact index). See/hear/try + downloadable artifact.
• Speech-in-noise — adaptive digits-in-noise (TTS + Web Audio noise) → relative SRT. Hear/try.
• Research tests — frequency discrimination (cents/Hz), gap detection (with catch trials/false alarms), temporal-modulation detection (4/8/16/32 Hz), binaural ILD/ITD lateralisation, dichotic stream attention, tinnitus spectrum/likeness map, residual-inhibition function, tinnitus tuning curve. Try; results in JSON export.
• Measurement spine — safety gate, capability/quality context, momentary check-ins + trends, auditable event stream, emergency stop, Session Trace (JSON/CSV export, merge-import, per-record + full-history deletion). See/try.
• Sound Lab + prospective A/B — bounded comfort sessions and computer-randomised within-person A/B (descriptive difference only). Hear/try.
• Result delivery (authored UI + optional dependency) — on-screen results, JSON/CSV export, audiogram PNG, pre-filled email; optional server-side send via Netlify Function → Microsoft Graph (the only third-party integration).

5 · Technical foundations & integrations

• Web Audio API throughout — OscillatorNode tones, BufferSource generated/filtered noise, StereoPannerNode L/R split, per-voice GainNode; canvas for audiogram/scope/spectrum. Exponential sweep law f(t)=f0·(f1/f0)^(t/T).
• Audiometry engine — ZEST-inspired posterior over θ (initPrior/updatePost/postStats), posterior-mean placement, SD-based stop; catch trials; shuffled interleave.
• Speech-in-noise math (verified in-repo against DiN literature): SNR = 20·log10(RMS_speech/RMS_noise); noise amplitude scaled by 10^(−N/20) ("fix speech, vary noise"); 4 dB→2 dB step after first reversal; SRT = mean of reversal SNRs (first discarded).
• No bundled local packages / no npm in the web app. Optional functions use Node 18+ global fetch (zero deps). Local packages named in the brief (FIRFilters/Kalman/SpectralProcessor/RemezEngine) are not used by this app — they belong to the sibling native/DSP portfolio; stating otherwise would be inaccurate.
• Integrations: optional Netlify Forms + Netlify Functions (send-result.js app-only client-credentials; send-result-delegated.js delegated/no-secret) → Microsoft Graph Mail.Send. Credentials live only in Netlify env vars; the recipient is fixed server-side (not an open relay); origin allow-listing, size cap, per-IP rate limit, honeypot, payload validation. tools/get-refresh-token.js mints the delegated token locally.
• Privacy posture: no microphone; no claimed room-acoustics/SPL/autonomic sensing; local-by-default storage (schema version 4); nothing transmitted without explicit user action + health-data consent.

6 · Maturity by feature (dated; evidence .git/logs/HEAD)

Dates derived from the git reflog (±1 day for timezone, commits +0100).

• Implemented + committed + public (source-complete): Quick, Audiology (Bayesian thresholds, audiogram, reliability), Speech-in-noise, measurement-quality layer — initial commit ~2026-08 (e12dfbd), Pages site + privacy/social-preview same window (8c2ea9d, 192cf6a).
• Research-test suite — "Publish research auditory tests" ~2026-09-14 (7bfe34e, codex/publish-research-tests, fast-forward-merged to main).
• Measurement spine + Sound Lab — ~2026-10-03 (1dd4032/7822f8e); private-by-default + master-bus monitor (b05ec98); prospective Sound Lab A/B (ecf7216); research-test accessibility + safety-lock polish ~2026-10-04 (edafbb7,a9a2a39,3e0a6c7). Last commit ~2026-10-04 (~13 days before this pass).
• Tests passed: No automated test suite in the repo. The README cites the author's own simulations (Bayesian ~6 trials/freq, ~1.8 dB mean error; SNR steps exact; SRT recovered to ~0.8 dB). Treat as author-stated, not independently verified.
• Built: N/A (static; no build).
• Runtime verified: the four earliest research tests (spectrum, residual-inhibition function, tuning curve, gap detection) were driven in-browser during their original build session; the later-added modules (measurement spine, Sound Lab, prospective A/B, and the additional research tests) are committed but NOT independently runtime-verified in this read-only pass — mark unknown.
• Released/deployed: deployment is configured . Live status not verified here (no network in this pass).

9 · Roadmap (aspirational — not yet built)

• Native iPhone HRV/pupillometry reactivity probe (sits with iOS Vitals) to add an objective autonomic correlate to the self-reported sound-reactivity log — the in-app text already states the browser does not measure this.
• Deeper tinnitus-tuning-curve vs psychoacoustic-tuning-curve comparison; richer longitudinal trend visualisations.
• Optional calibrated-headphone profiles to narrow the uncalibrated gap (still not clinical dB HL).

10 · Known limitations

• Uncalibrated: no dB HL/SPL; levels depend on device/DAC/OS volume/headphones; results are relative/within-session, not clinical.
• Headphones required for per-ear work; EHF (10–16 kHz) is hardware-limited and flagged exploratory.
• Not diagnostic; tinnitus "type" is a rule-based suggestion; research tests are within-session, uncalibrated, not clinical cut-offs.
• No automated test suite; several recent modules not independently runtime-verified here.

11 · FAQs

• Is this a medical device / diagnosis? No — an uncalibrated research/self-observation prototype; see an audiologist (calibrated audiogram) or ENT for concerns.
• Do I need anything installed? No — open the URL in a modern browser; headphones strongly recommended (required for per-ear tests).
• Where do my results go? Local to your browser by default; nothing is sent on finishing a test. You may export JSON/CSV, download the audiogram PNG, email, or opt into the configured send service.
• Why "relative" not dB HL? A browser can't produce calibrated sound pressure; the app re-references to your own best threshold and is explicit about it.
• Does it treat tinnitus? No. The Sound Lab offers bounded comfort sessions and a descriptive personal A/B; sound-therapy claims were tempered after a 2026 RCT — no efficacy is claimed.

12 · Research prototype vs clinical validation (required separation)

• Prototype capability (real): repeatable within-person relationships — adaptive thresholds with uncertainty, pitch/loudness/masking/residual-inhibition characterisation, speech-in-noise SRT, and the pitch-vs-loss geometry — all self-referenced and local.
• Clinical validation (explicitly absent): no calibrated dB HL, no booth/transducer calibration, no validated questionnaires (the impact index is original wording "modelled on" validated instruments), no automated/independent test suite, no regulatory status. The product's own copy states this throughout.

13 · Supporting "notched-sound research" (entry 13, repo = none)

Entry 13 ("research") has no separate repository; it is the embedded evidence strand that shaped the panel's notched-noise/Sound-Lab and tinnitus features — a literature review (via Consensus) cited inline in README.md and the Methods page: Bayesian audiometry [13–15]; Digits-in-Noise [4–6]; extended highs [8,9]; tinnitus battery/impact/dB SL/MRIL [10–12]; and the notched-sound-therapy evidence that was deliberately tempered — a 2026 RCT finding notched therapy no better than amplification [16], with tailored notched-music residual-inhibition effects [17,18]. Conclusion: this is supporting research embedded in the hearing panel, not a separate shippable product. The standalone notched-sound therapy app (NoiseColour) is a different deliverable owned by another agent (requests slug noisecolour, entry 14) and is out of scope here.