Built for regulated industries

Dictation that doesn't leave the machine.

DictateSI is a native desktop voice dictation tool built for banks, insurance, law, and real estate, where data has to stay local. Transcription and AI cleanup both run on your machine. No cloud, no API key. Deployable air-gapped.

Native app · an 8.7 MB download · No SaaS account required

Compose · Loan File 240518-Q3

To: J. Marsh, Underwriting

Re: Site visit notes, 412 Walnut

Property inspected at 09:14 on Monday. Roof in good repair, no visible settling on the south foundation. The applicant's stated use matches the structure: single-family with a detached garage converted to office.

Income verification on file

DictateSI is listening
Why DictateSI

A dictation tool your compliance team won't object to.

We built DictateSI because the dominant dictation product sends audio to the cloud, and the people we care about can't deploy it. DictateSI is engineered around the constraint instead.

On-device

Audio stays on the machine.

Transcription runs locally via Whisper. Audio is never sent to AWS, OpenAI, or any third party. There's nothing to subpoena off our servers because we don't have them.

On-device AI cleanup

AI polish that never leaves your machine.

Grammar and tone cleanup runs on a small model DictateSI downloads once and runs on your machine: no key, no cloud, works offline. The whole pipeline stays local. Prefer a cloud model? Connect the Anthropic, OpenAI or Azure account you already have. Either way, audio never leaves the machine.

Native

An 8.7 MB download, signed, MDM-deployable.

Real Swift, not Electron. Apple Developer-signed binary. Air-gapped install is supported. Your IT team can ship DictateSI with the same tools they ship the rest of your fleet.

How it works

Three steps. No SaaS account.

  1. 01

    Hold a hotkey, speak.

    DictateSI lives in your menu bar. Hold ⌃⌥Space anywhere in macOS (your email, a CRM, a spreadsheet) and dictate.

  2. 02

    Whisper transcribes, locally.

    Audio goes to whisper.cpp, running on the local CPU and GPU. Nothing leaves your machine. Latency stays under a second.

  3. 03

    Polished on-device, pasted at your cursor.

    A small AI model running on your machine cleans up filler words and punctuation: no key, no cloud, works offline. Say "number one, two, three" and you get a numbered list; "new paragraph" does what it says. (Prefer the AI account you already have? Connect it instead.) Domain vocabularies handle loan numbers, policy IDs, statute citations.

Built for

Teams that type the same paragraphs every day.

DictateSI is sold direct to mid-market employers. We start in Ohio and the broader Midwest, where we can put a person on the ground with your team during rollout.

Regional banks & credit unions

Loan files, faster.

Loan officers and processors type the same paragraphs every day: borrower interviews, site visit notes, decision memos. DictateSI turns 25-minute dictation into a clean draft in 4.

  • Loan officers
  • Processors
  • Branch managers

Insurance

Claims notes without the SOC2 dance.

Adjusters dictate field reports, call summaries, and claim updates. DictateSI stays inside the carrier's environment, with no third-party data processor to vendor-vet.

  • Claims adjusters
  • Customer service reps
  • Producers

Mid-sized law firms

Dictation, modernized.

Your attorneys already dictate. DictateSI replaces the cassette workflow with a tool partners trust: confidential, on-device, signed by you, deployed by your IT.

  • Paralegals
  • Attorneys
  • Legal assistants
DictateSI for law firms →

Real estate & title

Notes between showings.

Agents dictate listing notes, showing recaps, and settlement summaries from the car. DictateSI handles the transcription; your CRM gets the clean text.

  • Agents
  • Brokers
  • Title closers
The arithmetic

What does typing cost your team?

Most professionals type about 40 words a minute and speak about 120. The same borrower memo, claim note, or client letter takes roughly a third of the time by voice. Across a department, that compounds fast. Run your own numbers:

Your team

25
60

Notes, memos, emails, documentation. Not meetings, not reading.

$45

Salary plus benefits and overhead, per hour.

3.3 hrs

saved per person, per week

4,000 hrs

back across the team, per year

$180,000

of staff time, per year

How we compute this. Dictating at ~120 wpm produces the same text as typing at ~40 wpm in a third of the time, so each typing minute becomes roughly 40 seconds saved. Weekly figures assume 5 working days; yearly figures assume 48 working weeks. These are estimates to size the opportunity, not a promise. Adjust the sliders to your own numbers, and when our first pilots produce measured results, those will replace this arithmetic.

Deployment

Three ways to roll DictateSI out.

Pick the deployment model your security review will sign off on. We work with you and your IT team to make the rollout boring. That's the goal.

Direct download

Pilots

Apple Developer-signed binary. Drag to Applications. Use during evaluation or for solo practitioners.

MDM-deployable

Managed preferences

Ship the signed app via Jamf, Kandji, Intune, or whatever your fleet uses. The stable signature keeps permission grants across updates, and a standard configuration profile pins cleanup to on-device, allow-lists backends, forces the audit log, or locks an approved endpoint. Sample profiles are in the deployment kit.

Air-gapped on-prem

Available now

DictateSI runs with zero network access: pre-position the checksummed model files, keep on-device cleanup, and the machine never calls out. No telemetry, no update calls.

Deployment kit for your IT team: on-device-only profile, approved-endpoint profile, plain preference file, and the key reference with Jamf, Kandji and Intune steps.

Download

Get DictateSI for Mac.

One download. No account. Free while we're in beta; pricing arrives with v1.0.

DictateSI is a single native app: no SaaS dashboard, no cloud account. Drag it into /Applications, grant Microphone and Accessibility on first launch, and press your hotkey to dictate.

Public beta · cryptographically signed by the developer · Apple notarization arrives with the 1.0 release

First launch: one-time step

Because this beta isn’t yet notarized with Apple, macOS will say it “could not verify” DictateSI the first time you open it. That’s expected:

  • • macOS 15 or later: open DictateSI once, then go to System Settings → Privacy & Security, scroll down, and click Open Anyway.
  • • Older macOS: right-click DictateSI.app → Open → Open.

You approve once per machine. You can verify the signature yourself: codesign -dv /Applications/DictateSI.app

Highlights

Full changelog →
  • • Tap twice to keep recording: with a hold hotkey such as Fn, a quick double-tap keeps the recording open with nothing held; press once more to finish
  • • License keys, verified on the machine: no account, no activation server, no network. Free during the beta; a key changes nothing yet
  • • Only licensed components ship: every embedded library and downloadable model is MIT or Apache-2.0, stated exactly and guarded by a build test
  • • Your clipboard comes back: pasting a dictation no longer replaces what you had copied
  • • DictateSI's own cleanup model: the on-device model is now trained for dictation and measured against a written gate; existing installs keep working while it downloads
  • • Whisper large-v3-turbo: the most accurate on-device speech model, now the default on 16 GB+ Apple Silicon machines. Downloads in the background while you keep dictating
  • • Content guard: cleanup output that drops too much of what you said is discarded and your words are kept
  • • Silence pastes nothing: holding the hotkey without speaking inserts nothing, instead of a stray phrase
  • • Say the structure, get the structure: "number one, two, three" pastes as a numbered list, "bullet point" makes bullets, "new paragraph" breaks paragraphs, on every backend
  • • Managed preferences: IT can pin cleanup to on-device, force the audit log, or lock an approved endpoint with a standard configuration profile
  • • Connect the AI account you already have: Anthropic, OpenAI, Azure OpenAI, Gemini, Groq, Mistral, or your own gateway, as an opt-in alternative to the on-device default
  • • Tamper-evident audit log (off by default): a hash-chained, metadata-only record of whether anything left your machine. Export to JSON/CSV
  • • Verified model downloads: AI models are pinned to immutable revisions and SHA-256-checked before install
  • • On-device AI cleanup by default: no key, no cloud, works offline; guided setup on first run
  • • Read Aloud: select text anywhere, press ⌃⌥R, hear it read back on-device

System

Version
0.9.0 · beta
Size
8.7 MB
Minimum macOS
macOS 14 (Sonoma) or later
Memory
8 GB min · 16 GB recommended
Architecture
Apple Silicon (M1+) and Intel
Language
English (more planned)
Audio
Always on-device
AI cleanup
On-device · optional cloud

After download, drag DictateSI.app to your /Applications folder, then approve the one-time security prompt (see “First launch” instructions). Grant Microphone and Accessibility when DictateSI asks, and you’re dictating.

The app itself is tiny. Speech and on-device AI cleanup each fetch their model once on first use (Whisper 0.15 to 0.55 GB depending on your machine, the cleanup model ~1.1 GB) and cache it locally; after that, everything runs offline. Prefer not to download a model? Use the optional cloud or Ollama backend instead.

Deploying DictateSI across a team or need an air-gapped install? Book a pilot and we'll set it up with you in person.

FAQ

Honest answers.

If your question isn't here, ask it directly. We'll answer the security and compliance questions in writing.

Where does the audio go?

Nowhere. DictateSI records audio in memory, hands a 16 kHz PCM buffer to whisper.cpp running on-device, and discards the buffer. We don't write audio to disk by default, and there's no upload path in the binary.

What about the AI cleanup step?

Cleanup runs on-device by default: a small open-licensed model (Qwen2.5, Apache-2.0) runs on your machine via embedded llama.cpp, so nothing is transmitted and it works offline. If you would rather use an AI account you already have, connect it: Anthropic with your own key, or OpenAI, Azure OpenAI, Gemini, Groq, Mistral and any OpenAI-compatible gateway your IT team runs. Only the transcribed text, never the audio, goes to that provider under your account, and the audit log records every such send. Either way, audio never leaves the machine.

How is this different from Wispr Flow?

DictateSI keeps everything on the machine: transcription and AI cleanup both run locally, and there is no upload path in the binary. Cloud dictation tools work the other way; Wispr Flow's own privacy page states that transcription always happens in their cloud. If your data can't leave the device, that difference is the whole decision. DictateSI is also a native app, an 8.6 MB download, and needs no account.

Can our IT team deploy it via Jamf / Kandji / Intune?

Yes. The binary is Apple Developer-signed with a stable signature, so TCC grants (microphone, accessibility) persist across updates. Managed preferences work the standard way: push a configuration profile in the com.dictatesi.mac domain to pin cleanup to on-device, allow-list which backends people may pick, force the audit log on, turn local history off, or lock an approved endpoint. Users see "Managed by your organization" on locked controls. Sample profiles and the key reference are in the deployment kit linked on the security page.

Do you support Windows?

DictateSI is Mac-first today. A Windows version that works the same way, with all processing on the machine, is planned: the speech and AI engines DictateSI is built on already run on Windows, so the privacy architecture carries over. If your fleet is mostly Windows, say so in the pilot form below; demand directly sets the timeline.

How do we prove to an auditor that nothing left the machine?

Turn on the audit log (Settings → Audit Log, off by default). Every dictation appends a hash-chained, metadata-only record: which cleanup backend ran and whether anything left the machine, never the transcript. Export it to JSON or CSV and hand over the headline figure: off-device events: 0. Tampering with the log breaks the chain, and the built-in integrity check says so.

What does it cost?

Free for individuals during the beta. Teams start with a founder-led pilot, and per-seat pricing is set together from what the pilot shows. See the pricing page for the full breakdown.

When can we pilot?

We're onboarding the first cohort of pilot customers across Ohio and the Midwest now. Book a call below.

Talk to us

Pilot DictateSI with your team.

We onboard pilot customers personally: we'll come to you, install on a couple of seats, train your team, and stay involved through the first month.

Based in Ohio. Working across the Midwest first.

Prefer email? Write to aamir.khan98@outlook.com