← product

JQ Speech Mirror — therapist-recommended self-awareness on the glasses

Status: strategy / prep. COUNSEL-GATED. Nothing patient-facing is built. Authored 2026-07-11. A clinical-adjacent direction for Downstream Glasses (Mentra Live): a patient, on their therapist's recommendation, opts into a small list of phrases or speech patterns they want to catch themselves using — and the glasses reflect them back, gently, in the moment they're said. This doc sets the bright lines first; the only thing built so far is a generic, non-clinical detection primitive that ships off by default.

Inherits the Living Voice ethics wall and the owned-hardware on-device-first spec; shares the counsel gate of Legacy Channel eldercare.

Thesis

A patient wants to notice when they say things like "I always mess this up," "it's hopeless," "I'm such an idiot," or a personal verbal tic. Noticing is the whole intervention — but by the time they journal it that evening, the moment's gone. The Speech Mirror is the glasses catching it at the point of speech: the patient's own mic hears the phrase they chose to watch for, and a gentle, non-punitive cue lands in the moment. Over weeks, the catches get earlier and rarer. That's the point.

Why glasses (and not a phone/app): the catch has to be in the moment, hands-free, and private. A phone can't listen ambiently; a therapist isn't in the room; a journal is hours late. The glasses are there, on-device, when the words are actually spoken.

What it is NOT — the bright lines (load-bearing)

These are not caveats; they are the design.

  1. Not therapy, not diagnosis, not treatment, not a medical device. It's a mirror the patient chose to hold up. Positioned as a between-sessions awareness aid a therapist may recommend — like recommending a journaling habit — nothing more.
  2. The mirror reflects; it never coaches or judges. No "you're catastrophizing again." The cue is neutral and non-punitive — a soft tone, and/or a private count the patient reviews later — and the patient chooses it. (Inherits Odessa's capture-for-the-writing-never-live-coaching spirit: reflect, don't advise.)
  3. Therapist-recommended, not therapist-operated or therapist-surveilled. The patient configures it, owns the data, and it lives on their device. Any patient→therapist sharing is a separate, heavily-consented, HIPAA/counsel-gated layer — never the default, possibly never built.
  4. Not always-on surveillance. Opt-in; on-device matching (the utterance is matched locally and never sent anywhere); only the counts/notes the patient chooses to keep are kept, and they are deletable; an instant pause is always one tap away.

Architecture — reuses the shipped glasses loop

  • session.events.onTranscription already surfaces the wearer's speech from the Mentra Live mic (apps/glasses/src/session.ts).
  • The Speech Mirror is a watch layer: match each utterance against the patient's watchlist locally (the same decide-locally principle as Sophia — the utterance never leaves the device), and on a match fire a debounced gentle cue so it can never nag.
  • Primitive: apps/glasses/src/watch.ts — pure detection (checkUtterance) + a cooldown-debounced createWatcher. Off by default (WATCH_ENABLED, unset) and fails safe, exactly like Sophia. Wired into the session behind that gate; inert until the clinical product + consent exist.
  • The cue is neutral by design: a soft configured tone (WATCH_CUE_AUDIO_URL) and/or a silent increment — never a spoken judgment.

Ethics / clinical guardrails (before anything patient-facing ships)

  • Counsel gate. Clinical-adjacent, potentially HIPAA the moment a therapist touches patient data. No patient- or therapist-facing surface ships without counsel review — same gate as Legacy Channel.
  • Consent + ownership. The patient opts in, builds their own watchlist, and can delete everything. They're told plainly: a self-awareness aid, not treatment.
  • On-device-first. Utterance matched locally; nothing leaves the device but patient-chosen aggregates. Inherits the owned-hardware ethics spec.
  • Non-punitive, patient-controlled. Neutral cue, debounced, adjustable sensitivity, instant pause. A tool that shames is a tool that harms — the cue must read as a friend's quiet "hey," never a buzzer. This is the load-bearing UX rule.
  • Clinician + counsel review of all copy/marketing; explicit "not diagnosis/treatment/medical-device" framing.

Phased plan

  1. Speech-watch primitive (built, non-clinical, OFF by default). watch.ts detection + gentle-cue hook + tests, WATCH_ENABLED-gated. Reusable, hardware-ready, nothing patient-facing. ← this slice
  2. Patient watchlist config + private review (counts over time). Clinical consent flow. Counsel-gated.
  3. Therapist-recommendation packaging — a "recommend Downstream Glasses to a patient" flow for therapists (recommend, not operate). Clinician + counsel review of every word.
  4. (Only if warranted) opt-in patient→therapist aggregate sharing — heavy consent, BAA/HIPAA, counsel-gated, default off. A separate product decision, not assumed.

See also

JQ SPEECH MIRROR — Docs | HiveJournal