What changed
FACT (source id 6998 / get-inscribe benchmark): Apple shipped a first-party on-device SpeechAnalyzer API for macOS/iOS with accuracy benchmarked against Whisper. FACT (Show HN, Biscotti): a shipping macOS app already does on-device, no-bot meeting transcription — proving the local-only pattern is buildable today.
Why now
On-device high-accuracy STT removes the one thing that made confidential-audio transcription legally awkward: network egress to a third-party STT vendor requiring a BAA. With inference on the device, the compliance objection and the per-minute cloud bill both disappear at once.
Converging signals
Two capability signals meet: (1) Apple first-party on-device SpeechAnalyzer (new, free, no GPU/cloud bill), and (2) an existing local-only meeting-AI app corroborating the architecture is viable. This is a capability×capability convergence, NOT a mandate — there is no forced-filer class here.
Customer pain
HYPOTHESIS (not evidenced in input — demand_evidence is EMPTY): solo therapists spend unpaid hours writing progress notes and either avoid cloud scribes over privacy fear or pay $99-150/mo for them. The pain is real in the category but NONE of it is cited in this input; treat as unverified until pain/spend evidence is pulled.
Who pays
Solo therapists, psychologists, counselors; secondarily solo attorneys wanting offline session/deposition notes. Discretionary buyer paying by card monthly — not a forced buyer.
Solved today
Cloud AI scribes with BAAs (Freed, Nabla Copilot, Upheal, Mentalyc, Heidi), generic dictation (Otter, Dragon), or typing notes by hand. On-device: Biscotti already exists for meetings.
Why current solutions are bad
Cloud scribes require trusting a vendor with the most sensitive audio that exists, carry a monthly floor, and still leave a BAA/paper trail. Hand-typing is the real time sink. The on-device angle genuinely differentiates on privacy AND cost — but only on those two axes.
Proposed product
Sandboxed native macOS/iOS app, network entitlement stripped, that records a session, runs SpeechAnalyzer locally, maps the transcript to a vertical note template (SOAP / DAP / progress note), and exports PDF or into an EHR. Sell the 'nothing leaves your Mac' guarantee as the wedge.
MVP version
Single-vertical (mental-health progress notes) macOS app: record → on-device transcript → one editable SOAP template → PDF export. Ship the accuracy KILL TEST first: 10 clinicians, real session audio with jargon/accents, measure edit-time saved vs typing. If it fails that, stop.
30-day build
Build recording + SpeechAnalyzer pipeline + one note template + PDF export. Run the 10-clinician accuracy test on real jargon-heavy audio. Kill or continue based on measured edit-time savings.
60-day build
If accuracy passes: add a second template (DAP), speaker separation, and a simple local library. Recruit 20-30 beta clinicians via therapist subreddits/Facebook groups and privacy-forward positioning; App Store submission.
90-day revenue plan
Convert betas to a $29-49/mo subscription (undercut $99-150 cloud incumbents on price + privacy). Realistic: dozens of paying users, low-thousands MRR — a modest niche product, not a breakout.
Distribution path
App Store + direct download; content/SEO around 'HIPAA on-device therapy notes,' therapist communities, privacy angle. Distribution is HARD: reaching thousands of scattered solo clinicians is slow and semi ad-dependent — the weakest part of this idea.
Pricing hypothesis
$29-49/mo flat (one device, unlimited local sessions). No per-minute cost floor is the pricing advantage.
Technical difficulty
Moderate. Native Swift app + audio capture + template engine + EHR export mappings. No cloud infra, but real app-engineering and accuracy tuning on domain jargon/accents (the KILL TEST is the whole ballgame).
Legal / regulatory risk
Marketing 'HIPAA-safe' is a claim to make carefully — on-device reduces but doesn't eliminate the clinician's own HIPAA obligations, and overclaiming is a liability. Not licensure-blocking for the founder, but requires disciplined copy.
Platform dependency
HIGH and structural: the entire moat is Apple's SpeechAnalyzer + App Store. Apple bundling equivalent on-device dictation into Notes/Health (input estimates ~9 months) collapses the wedge. Also App Store review risk on health-adjacent apps.
Founder fit
BELOW his sweet spot. This is a native consumer/prosumer app in a crowded AI-scribe market — NOT a government-portal forced-filer play (his proven edge and highest-fit shape). No public money, no mandate, no per-filing hook. Swift app-building and B2C therapist distribution are outside his stated strengths (public records, compliance monitors, industrial/data tools).
Breakout potential
Low-to-moderate. Real niche, but capped by a short platform-defined edge, entrenched well-funded incumbents, and a hard-to-reach fragmented buyer.
Final recommendation
PASS / REVISIT-only for this founder. The on-device angle is genuinely clever and the category has money in it, but it's a crowded, platform-dependent B2C app far from his forced-filer edge, with no demand evidence in-hand and a short Apple-defined window. If pursued at all, do it ONLY as a fast, cheap KILL-TEST spike (accuracy on real clinical audio) before any real build — and prefer redirecting the same energy toward a public-money/compliance-filing opportunity that fits him.
Next action
Run the accuracy KILL TEST cheaply first: feed 30-60 min of real jargon-heavy therapy session audio through Apple SpeechAnalyzer and measure word error rate + edit-time-saved vs typing. If it clearly beats typing, pull real pain/spend evidence from therapist communities before committing; if not, drop it.