How Signal turns a 50-minute session into a finished, signed note.
No marketing fluff. This page walks through exactly what happens from the moment the phone rings to the moment the insurance payment lands — the technology behind it, the safeguards around it, and the honest limitations of it.
What you get, in numbers
8 sec
voice → emotion signal on screen
8 min
average note time, draft to signed
94%
first-pass insurance acceptance
Zero
raw audio retained
01
Before the session
Amelia has the chart ready before they walk in.
The work that used to live in your evenings — chasing forms, running eligibility, holding copays, scanning intakes for risk language — happens in the background, automatically. By the time the session starts, the chart is current, the benefits are verified, and you know what you're walking into.
Intake forms sent the moment the slot is booked
Demographics, history, presenting concerns, ROI, GFE acknowledgement. Smart skip logic — clients only see what applies. Most return them in under nine minutes.
Insurance verified in real time
Stedi runs a 270/271 eligibility check the moment a card is added. Copay, deductible, mental-health benefit limits, prior-auth requirements — all on the chart before you walk into the room.
Card on file authorized, not charged
Stripe pre-auth holds the copay so no-show fees become recoverable. The actual charge only fires after the session is signed.
Risk + readiness flags surface to you
Amelia reads the intake free-text fields and flags suicidal ideation language, recent medication changes, and prior trauma history — so nothing important hides on page 14 of the form.
Amelia handling intake forms, insurance eligibility, and payment authorization over chat — before the session begins.
The live affect view — gentle waveform on the periphery, no popups, no pings. You watch the client, not the screen.
~280ms
Transcript latency
~8s
Affect signal latency
0 KB
Raw audio stored
02
During the session
A second pair of senses — quietly, on the edge of the room.
Voice streams to Deepgram for transcription, then to our own affect engine that turns vocal cues into a regulation timeline you can glance at. No raw audio is ever stored. Crisis flags fire fast and quietly. The room stays the room.
Voice → emotion in roughly eight seconds
Deepgram streams transcription with diarization. Our affect engine layers valence, arousal, regulation, and cognitive load over the transcript — per speaker, with confidence intervals.
A regulation timeline you can glance at, not stare at
The sidebar lives at the edge of your peripheral vision. A single calm waveform that brightens when affect shifts. No popups, no pings, no tabs that pull your attention out of the room.
Crisis flags that fire fast and quietly
SI/HI language, dissociative cues, agitation ramps — surfaced as a small amber dot. You decide whether to engage. Signal never speaks in the room.
Per-speaker tracks for couples and family
Two waveforms instead of one. Pursue–withdraw, criticize–defend, and other dyadic cycles surface as gentle pattern markers — useful with EFT, Gottman, and IFS work.
Couples & family work
Two waveforms. One cycle picture.
Pyannote diarization gives every speaker their own affect track, so pursue–withdraw, criticize–defend, and other dyadic cycles surface as gentle pattern markers. Useful with EFT, Gottman, IFS, and any modality that cares about the dance between two people.
- Per-partner valence and arousal tracks
- Cycle markers for common relational patterns
- Couples-specific note formats included
03
After the session
A finished, signed note in eight minutes — not eight at night.
The session ends. You pick a format and a modality. Signal drafts a note grounded in the actual transcript and affect events — never invented. You read it, edit anything you want, sign it. The medical-necessity language for insurance is already there.
Pick a format, pick a modality
SOAP, DAP, BIRP, GIRP, PIRP, SIRP, PIE, or your own template. Pair it with one of 16 modalities — CBT, ACT, DBT, EMDR, IFS, EFT, Gottman, MI, Schema, Narrative, and more — so the clinical voice matches what you actually did.
Draft grounded in transcript and affect, not invented
Every assertion in the draft links to a transcript moment or a regulation event. Hover any line to see the source. No paraphrased fabrication — if Signal can't ground a claim, it doesn't make the claim.
Medical-necessity language for insurance, automatically
PHQ-9, GAD-7, PCL-5, and other measures auto-link into the medical-necessity paragraph. ICD-10 codes suggested from content; you confirm. No more rewriting the same justification you wrote yesterday.
You sign. Signal locks it.
Once you sign, the note becomes immutable in the audit log — addenda only after that point. The signed PDF is generated, attached to the chart, and pushed to the claim queue for Amelia.
Eight clinical note formats, sixteen modalities. You pick; Signal drafts in your voice and your style.
Amelia builds the 837P, scores denial probability, submits through Stedi or TELUS, and reconciles the ERA when it lands.
04
Behind the scenes
The note becomes a claim. The claim becomes payment.
Once you sign, Amelia takes over. She builds the claim, predicts whether it will pay, submits through Stedi (US) or TELUS (CA), reconciles the ERA when it returns, and learns from every outcome. Your A/R aging report stops being a horror story.
Amelia builds the 837P claim
She pulls CPT, ICD-10, place-of-service, and modifier set from the signed note. She knows your payer-specific quirks — Aetna's preferred POS, Cigna's modifier conventions, Blue Cross regional fork — because she's seen them paid before.
Predicted denial probability before submission
Each claim gets a score. If it's above a threshold, Amelia holds it and tells you what's likely to bounce — wrong modifier, missing prior-auth, units mismatch — so you can fix it before submitting, not after.
ERA auto-posting and denial autopilot
835s come back through Stedi, post against the claim, and reconcile against the patient ledger. Denials get categorized, queued, and auto-resolved when the fix is something Amelia has learned (which, after a few weeks, is most of them).
She gets sharper every week
Every paid claim, every denial, every appeal outcome trains the per-payer rules. By month three, your first-pass acceptance rate is materially higher than it was on day one.
Under the hood
We picked boring infrastructure on purpose.
Signal isn't built on novel infrastructure or experimental models. It's built on the boring, reliable, audit-friendly stuff — Google Cloud, Stripe, Stedi, Deepgram, Anthropic — wired together with care and a lot of tests. Boring is a feature when the data is PHI.
798
API endpoints
262
automated tests
47+
frontend pages
5
regions covered (US + CA)
The stack, layer by layer
Speech-to-text
Deepgram Nova
Real-time streaming with sub-300ms latency. HIPAA BAA in place.
Speaker diarization
Pyannote.audio
Open-source, self-hosted. Per-speaker tracks for couples, family, and group.
Affect & regulation engine
Custom (Signal)
Built in-house. Five-dimension model — valence, arousal, regulation, load, variability.
Clinical reasoning
Anthropic Claude
BAA-covered model calls. Grounded in transcript + affect, never invents quotes.
Database
Google Firestore
Multi-region. Per-document audit log on every PHI access.
Compute & infra
Google Cloud (us-central1)
Cloud Run for the API. Cloud Scheduler for background jobs. SOC 2-aligned.
Encryption
AES-256 in transit + at rest
TLS 1.3 on the wire. Customer-managed keys available on request.
US insurance
Stedi clearinghouse
3,400+ payers. Real-time 270/271 eligibility, 837P claims, 835 ERA.
Canada insurance
TELUS eClaims
29+ insurers, instant adjudication. Insurance receipts for everyone else.
Payments
Stripe
Pre-auth, payment plans, A/R aging. Cards never touch our servers.
Multi-tenant isolation
Per-clinic row-level security in Firestore. A practitioner in Clinic A cannot read Clinic B data — enforced at the database, not the app.
Region-aware data residency
Canadian clinics' PHI stays in Canadian regions. US clinics in US regions. Configurable per-tenant for enterprise rollouts.
The honest answers
Common worries, answered honestly.
These are the questions therapists actually ask in our pilot calls. We answer them the same way here as we do on those calls. No spin.
Will it disrupt the therapeutic alliance?
Honest answer: only if you let it. Signal never speaks, never beeps, and never opens a window in front of you. The affect timeline lives in your peripheral vision — many therapists glance at it twice a session and ignore it the rest of the time. Clients consent to recording up front; in our pilot cohort, fewer than 1% declined after the first session, and the ones who did get the same product without the affect track. Your eyes stay where they belong.
Is my client data safe?
Yes, and we're specific about what that means. We do not retain raw audio after the session ends — only encrypted transcripts and embeddings, and only with explicit, revocable client consent. Every PHI access is logged at the row level. We're HIPAA-compliant in the US (signed BAAs with every subprocessor that touches PHI), PIPEDA-compliant in Canada (with PHIPA, Law 25, HIA, and PIPA province controls), and our infrastructure is SOC 2-aligned. If you ever want your data out, you get a full export within 72 hours, no questions asked.
What if it gets a note wrong?
It will sometimes — that's true of any documentation tool, including the one you're using right now. The difference is how we handle it. Every line in a Signal-drafted note is grounded: hover any sentence and you see the transcript moment or affect event it came from. You read the draft, edit anything you want, and sign. Nothing leaves the system unsigned. If the draft is materially wrong, the edit teaches Signal — your style and your corrections feed your own private model so the next note is closer to what you'd have written.
Does Signal replace my clinical judgment?
No. Signal is decision support, not a decision-maker. The affect timeline shows you what the voice signal looks like; you decide what it means in the context of this client and this relationship. The note draft is exactly that — a draft. The treatment plan is yours. The diagnosis is yours. The risk assessment is yours. Signal's job is to remove the typing, not the thinking.
How long does setup actually take?
Most therapists are seeing clients on Signal inside 90 minutes. Practices migrating from SimplePractice, Jane, or TherapyNotes typically run parallel for a week (we import clients, appointments, and notes), then fully cut over inside 30 days. We do this with you on a shared call if you want, or you can self-serve from the migration center. There's no per-seat onboarding fee.
What about insurance claims that get denied?
Denials are categorized and queued automatically. Most categories — wrong modifier, missing referring provider NPI, units mismatch, eligibility lapsed — Amelia can resolve without human intervention because she's learned the fix from prior payers. The ones that need a human (medical necessity, appeal-with-records) are surfaced to your billing dashboard with a one-click drafting flow. Average days-in-AR drops materially in the first two months for most practices.
Can I try it without committing to a migration?
Yes. The 14-day free trial is the full product — clinical AI, AI notes, Amelia for chat and SMS, telehealth, billing, all of it. No credit card required. Many therapists run two or three real sessions in the trial just to see what the affect timeline does in their actual room with their actual clients before deciding anything.
You stay in control
Consent, off-switches, and exits — all real.
Every client signs a recording consent before audio is processed. They can revoke it at any time. You can disable the affect timeline per-client, per-session, or globally. Signal works as a regular EHR for any client who prefers it that way.
Per-session record toggle
Default off until consent is on file. Visible to the client and the clinician at all times during the session.
Revocable consent
Clients revoke from their portal. Revocation triggers immediate deletion of transcripts and embeddings tied to their record.
Affect off-switch
Disable the affect timeline per-client or globally. Notes still draft from transcript alone if you prefer.
72-hour data export
Full export of charts, notes, billing history, and audit logs in standard formats — within 72 hours of request.
Tuesday, 2:00 PM
One session, end to end, in real time.
A look at the choreography across roughly two hours, from the moment a returning client books to the moment the claim posts.
T-24 hrs
Reminder + intake refresh
Reminder fires. Amelia notices the PHQ-9 hasn't been updated in 30 days and asks the client to retake it. They do, in 90 seconds.
T-30 min
Eligibility + payment hold
Stedi runs a fresh 270/271 eligibility check. Stripe pre-authorizes the copay. Both results land on your dashboard before you walk in.
T-0
Session begins
Client consents to recording (on file from prior visit, confirmed today). Deepgram starts streaming. Affect engine warms up. The room is just the room.
T+12 min
Affect shift detected
Client moves into a regulation valley around work stress. The waveform on your sidebar dims. You notice. You stay with it.
T+50 min
Session ends
Audio buffer is cleared from memory. Transcript and affect events are encrypted and stored. Raw audio: zero bytes retained.
T+58 min
Note signed
You pick SOAP + CBT. Signal drafts a 380-word note grounded in the transcript. You read, edit two sentences, sign. Note is locked, PDF generated, copay captured.
T+59 min
Claim built and submitted
Amelia builds the 837P with CPT 90834, payer-specific modifier, and place-of-service. Denial probability scored at 4%. Claim auto-submits to Stedi.
T+72 hrs
ERA posts
835 returns, posts against the claim, reconciles to the patient ledger. Variance under threshold — no human review. Days-in-AR for this visit: 3.
See it in your own room with your own clients.
Fourteen days. Full product. No credit card. The best way to evaluate Signal is to run two real sessions in your trial — the affect timeline tells you everything in five minutes.
Questions first? Email hello@signalehr.com — a real human reads it.