Three tools keep your sessions on your own computer. Here is how they differ.
Scribular ($99 one-time), Confidant ($359 one-time or $29.99/month), and Spokenroom ($59 one-time) all write SOAP/DAP-style notes locally. They differ on price, on how they tell your voice from your client's in telehealth, on how strictly the note is checked against the transcript, and on which computers they run on. The table is the whole comparison; the rest of the page explains the rows.
Two other tools also keep your sessions on your own computer.
Here is how we differ — and when you should pick one of them instead. Cells where they are better are in bold. "Not stated" means the vendor's site did not say, not that the feature is absent.
| Spokenroom | Scribular | Confidant | |
|---|---|---|---|
| Price | $59 one-time (1 year of updates; $29/yr to keep updating, optional) | $99 one-time (labelled early-access pricing, stated to increase after launch) | $359 one-time (1 year of updates; $179/yr optional) or $29.99/month ($24.99/month billed annually) |
| Free trial | 7 days, no card, no account | 14 days, no card | 14 days, no card |
| Devices per license | 2 | 2 | not stated |
| Platforms | Windows 10/11 (x64, AVX2 CPU) · macOS in v1.5 | Mac + Windows | macOS 13+ Apple Silicon only · Windows 11 |
| Minimum memory | 8 GB (transcript + plain summary) · 16 GB for clinical templates | not stated | 8 GB (Lite model) · 16 GB recommended (Pro model) |
| Where audio is processed | Your computer | Your computer | Your computer |
| Telehealth recording | Yes — microphone + system audio captured as two separate tracks | Yes — captures Zoom / Meet / Teams / SimplePractice / VSee locally | not stated |
| Who-said-what in telehealth | Decided by track at record time — no acoustic guess, no model | Speaker tagging (therapist / client / custom names) | not stated |
| Transcription during the session | Yes — transcript is ready when you stop | Notes within the session or immediately after | Not stated; generation after recording: up to one minute (16 GB) / closer to four minutes (8 GB) |
| Note formats | SOAP · DAP · BIRP · GIRP · custom Markdown template | SOAP · DAP · BIRP · GIRP | SOAP · DAP · BIRP · GIRP · PIRP · SIRP · custom prompts |
| Guard against invented content | Two-stage: evidence first, then prose. Every claim cites a transcript line; quotes must be exact substrings; names, medications, dates and numbers must appear in the cited line; an empty field is written N/A by code, never by the model | not stated | Sentence-level check: any sentence whose words cannot be found in the recap is refused |
| Dictation mode (no recording) | v1.5 | Yes | not stated |
| Runtime | Native app + signed sidecar processes; no Python, no Ollama; installer ≤150 MB | whisper.cpp + bundled Ollama | not stated |
| Published benchmarks | Yes — raw CPU-only numbers, old and new hardware | No | No |
| Other verticals | General edition $39; legal edition v1.5 | Therapists only | Therapists only |
| Refunds | 30 days | not stated | 30 days |
| Who makes it | Small independent team | Independent developer, Seattle, WA | Founded by a licensed therapist (LCSW-R, 16 years in practice); self-funded |
Scribular and Confidant figures from their public websites, checked 2026-10-18. We recheck quarterly; if you find a row out of date, email us and we will fix it the same week.
Choose Confidant if
- You want a tool built by a practising therapist. The founder is an LCSW-R with 16 years in practice, and that shows in the template set (PIRP and SIRP included).
- You prefer a monthly plan and never want to think about update renewals.
- You are on an Apple Silicon Mac today and can't wait for our macOS build.
Choose Scribular if
- You want post-session dictation as a first-class mode right now (ours lands in v1.5).
- You are on a Mac today.
Choose Spokenroom if
- You run telehealth and have been burned by a sentence attributed to the wrong speaker. We record you and your client on separate tracks, so attribution is a fact of the recording, not a guess made afterwards.
- You want the note checked against the transcript by code, line by line, and an empty field to say N/A rather than something plausible.
- You want the transcript finished the moment the session ends, and to go straight into the next appointment while notes write themselves in a queue.
- You are on Windows 10, or an Intel Mac (v1.5), or a laptop with a modest CPU — and want to see the actual benchmark for hardware like yours before you pay.
- You also interview, research, or practise law, and want one engine with a template set for each.
- Price matters: $59 is the lowest of the three.
Where the differences come from.
1. Who said what in a telehealth session
In a video session your microphone carries only your voice and your computer's audio output carries only your client's. Spokenroom records those as two separate tracks and never mixes them. Speaker attribution is therefore decided at the moment of recording, with no model involved and nothing to get wrong. In-person sessions have one microphone, so there we do use acoustic speaker separation — and we show you the transcript and note first instead of making you wait for it.
2. How the note is constrained
Small local models are good at writing fluent prose and bad at leaving a field empty. In our own testing a 1.7B model filled an "Objective" section with six quotations that did not exist in the transcript. So Spokenroom splits note-writing in two. First the model extracts candidate facts, each with the transcript line it came from and the exact quote. Code then checks every one: the line must exist and belong to the right speaker; the quote must be an exact substring; names, medications, dates and numbers must appear in that line. Only facts that pass are handed to the second stage, which writes the field from them and nothing else. If no fact passes, code writes "N/A" — the model is never asked to fill a field it has no material for.
Confidant describes a related safeguard: refusing any sentence whose words cannot be found in the session recap. We think both approaches are the right instinct. Ours checks where each claim came from and who said it, not only whether the words appear.
3. We publish numbers from a slow machine
We develop on an eight-year-old CPU-only workstation on purpose, and we publish the raw timings from it and from a current laptop with the release. See requirements.