Case notes, dictated privately on your Mac
Audio stays on the laptop. Cloud never sees the recording.
Short answer: Voice to text for social workers on Mac is most useful when client PHI never leaves the laptop. MetaWhisp runs Whisper large-v3-turbo on the Apple Neural Engine — free, on-device, no account, audio never uploaded in local mode. The realistic workflow is dictating case notes at your Mac between home visits, then pasting them into your agency's EHR. Our only first-party accuracy figure is 2.76% WER on LibriSpeech test-clean; case-note terminology has not been benchmarked, so expect some cleanup.
Why does on-device voice-to-text matter for social workers?
Social work case notes are full of protected information: client names, dates of birth, addresses, diagnoses, household members, sometimes immigration status. Most "smart" dictation apps send that audio to a remote server for transcription. Some keep it, some anonymize it, some sell aggregated "training data" rights back to themselves in the terms of service. Whatever the fine print says, the audio has left your laptop and is now someone else's problem to secure.
That's the real differentiator for an on-device tool: the audio file never leaves the machine. No round trip, no server breach later, no second vendor in the chain of custody. For a profession that handles identifying information daily, that distinction matters more than accuracy benchmarks.
MetaWhisp's local mode runs Whisper large-v3-turbo through WhisperKit directly on the Apple Neural Engine. Audio is captured, transcribed on your Mac, and discarded. No telemetry, no upload. The app works fully offline once the ~950 MB model is downloaded — you can pull the Wi-Fi and dictate. Pro removes that constraint (it sends audio to MetaWhisp's cloud for transcription) and is the wrong tier if PHI containment is your priority. The relevant page is on-device transcription, where the architecture is described in plain language.
The realistic workflow: dictate after the visit, not during it
I want to set expectations honestly. MetaWhisp is a macOS desktop app. It doesn't run on your phone. You cannot stand in a client's living room and dictate into your iPhone with it. If your agency workflow depends on live field dictation — talking into a phone while walking out the door — this tool, as of today, is not built for that.
What it is built for is the other half of the day: pulling into the parking lot, getting back to your office Mac, opening the laptop, and unloading three home visits into a structured case note before lunch. That's the window most social workers actually have for paperwork. It's also the window where privacy risk peaks — you're tired, your handwriting is bad, and the temptation to dictate into a phone-based cloud app is real.
Pro tip: Keep a small notebook (paper, not the laptop) for the 5–10 facts you can't afford to forget on the drive back: client's preferred pronouns, the name of the partner who answered the door, the exact concern they raised. Dictation is faster than typing for the body of the note, but it's worse for the tiny high-stakes details if you're back at the laptop 90 minutes later.
Here's how the workflow looks in practice:
- Drive / commute back to the office. Jot the 3–5 anchor facts in your notebook.
- Sit at your Mac. Hold the global hotkey (default: Right Option ⌥), speak the case note in your normal cadence. MetaWhisp listens, transcribes locally, and copies the result to your clipboard.
- Paste into your EHR or word processor. Cmd-V into the chart. Edit the terms Whisper got wrong (almost always names and abbreviations).
- Optionally run an AI pass via processing modes to clean up punctuation, fix the odd tense, or restructure into SOAP format.
- Save and close. Audio is already gone — the local mode pipeline does not keep recordings.
How accurate is the dictation on social work terminology?
I have to be careful here because accuracy numbers are where articles like this go wrong. The only first-party accuracy number we have published is from a LibriSpeech test-clean run, which measured 2.76% WER (~97% accuracy) for our pipeline. That's a benchmark on audiobook recordings with clean audio — not on case-note dictation with street names, mental health terms, and kids' nicknames.
We have not benchmarked accuracy on social work terminology. Nobody has, to my knowledge, published a public first-party figure for that domain. Anyone claiming "98% accurate on case notes" or "the best for HIPAA" is making it up.
What you can reasonably expect from MetaWhisp on case notes: very strong performance on ordinary English prose, occasional mistakes on unusual names (especially non-English ones), and frequent struggles with acronyms and rare terminology. The "Correct" and "Rewrite" processing modes help most here — they take the raw transcript and apply your own OpenAI or Cerebras key (BYOK) to clean it up, with only the text (never the audio) leaving your machine. Anyone promising domain-specific accuracy for clinical or legal transcription without naming the test set is selling, not measuring. For the technical background on Whisper, the open-source model is documented on GitHub; the large-v3-turbo variant is what we ship.
How MetaWhisp compares to other tools for case notes
Most "voice to text for social workers" content I've seen compares the wrong things. The right question isn't which tool has the slickest iOS app — it's which tool can be in your lap on a Mac, in a county office, with audio that never leaves the building.
| Tool | Where audio goes | Free tier | Apple Silicon native | Phone field dictation |
|---|---|---|---|---|
| MetaWhisp (local mode) | Stays on your Mac | Unlimited, no account | Yes | No |
| MetaWhisp (Pro cloud mode) | Sent to MetaWhisp servers | — | Yes | No |
| Apple Dictation | Apple servers | Free | Yes | Yes (iOS) |
| Phone-based transcription apps | Varies by vendor | Varies | — | Yes |
Two things stand out. First, "free, on Apple Silicon, with audio that doesn't leave the machine" is a small club — most free tiers are free because you're the product. Second, none of these tools that I know of handle the live-in-the-field piece on a Mac. If your agency requires phone-based field dictation, you'll need a separate workflow for that, and the privacy story will be different.
Founder's note: I run my own 7-app head-to-head WER test on the same audio (2026): MetaWhisp landed at 3.7%, three competitors in the ~3.5% range, Apple Dictation 11–14%, Google Docs voice typing 5–7%. Those numbers come from one person's recording on one Mac and they should be treated as a sanity check, not an industry benchmark. Names, jargon, and accented speech will move every number on that chart.
What's the difference between MetaWhisp Free and Pro for case notes?
If PHI containment is your priority, the answer is short: stay on Free, local mode, and add your own API key if you want AI polish. Pro is the tier that uses MetaWhisp's built-in cloud AI and adds cloud transcription — by design, that sends your audio to our servers. For social work case notes, that's usually a non-starter.
| Tier | Cost | Transcription path | AI processing | Best for case notes? |
|---|---|---|---|---|
| Free (local) | $0 | On-device only | BYOK (your OpenAI / Cerebras key) | Yes |
| Pro | $30/yr or $7.77/mo | Local or cloud | Built-in cloud AI (no key needed) | No — data leaves Mac |
For specifics on what each tier includes and what the limits are, see the pricing page. For the deeper "how does HIPAA even apply to a tool like this" question, our HIPAA local dictation guide walks through it without overpromising.
Setting up MetaWhisp for social work case notes
The setup is genuinely short — no agency IT ticket required. MetaWhisp is a downloadable .app, not an MDM-managed SaaS. (If your agency does manage your Mac, you'll need their permission to install it; that's between you and them.)
- Download and install. Grab the app from the download page. Requires macOS 14+ on Apple Silicon (M1 or later).
- First launch downloads the model. Roughly 950 MB, one-time. After that, MetaWhisp works offline.
- Pick your hotkey. Default is Right Option. Anything that doesn't collide with your EHR's shortcuts works.
- Add a BYOK API key (optional). If you want "Correct" or "Rewrite" modes to clean up transcripts, paste in an OpenAI or Cerebras key. The audio never leaves your Mac; only the transcript text is sent to your provider.
- Test once before a real visit. Dictate a paragraph of anything into TextEdit. Confirm paste works. Confirm the hotkey doesn't fight anything.
Is MetaWhisp HIPAA-compatible? Yes — local mode is HIPAA-compatible in the way that matters for solo dictation: the recording never leaves the Mac, no third party sees the audio, and you control which (if any) AI provider receives the transcript text. Compliance itself belongs to the practice, not any single app — no vendor holds a "HIPAA certification" issued by HHS — but local mode offers the architectural fit that HIPAA-compatible workflows need. That is enough to slot into many agencies' policies; it is not, and cannot be, a substitute for your agency's own privacy review. The relevant HHS material is at hhs.gov/hipaa for the underlying rule.
What MetaWhisp won't do for you
An honest list, because pretending a tool is bigger than it is the surest way to waste a social worker's afternoon.
- No iPhone or iPad app. Field dictation on iOS is a planned 2026 feature. Today, it's Mac or nothing. If your workflow requires phone dictation, this isn't the tool.
- No speaker diarization. MetaWhisp does not currently label who said what. If your visit included a multi-party conversation and you need to attribute quotes, you still need to add the labels yourself.
- No semantic transcript search. You can't yet search across all your old case-note transcripts for "client mentioned housing." That feature is on the roadmap.
- No clinical accuracy guarantee. We've benchmarked LibriSpeech. We have not benchmarked mental-health terminology, custody acronyms, or your county's specific form language. Expect cleanup.
- No automated EHR integration. We copy to your clipboard. Pasting into your chart is still a human step, and that human step is the right place for a privacy review.
Pro tip: Treat MetaWhisp as a typing accelerator, not a charting system. The 90-second job — get the body of the note down while it's fresh — is what it does. The EHR's role-based access controls, audit logs, and co-signature flow stay where they belong: in the EHR.
Frequently asked questions
Does MetaWhisp upload client information to a server?
In local mode — the free default — audio and transcripts stay on your Mac. The Whisper model runs on the Apple Neural Engine and the recording is not transmitted. If you turn on the optional AI processing modes ("Correct" or "Rewrite") and add your own OpenAI or Cerebras API key, only the transcript text is sent to your provider, never the audio, and never to MetaWhisp. Pro adds a cloud transcription option that does send audio to MetaWhisp servers; if PHI containment is the priority, stick to Free.
Is MetaWhisp HIPAA-compatible?
MetaWhisp's local mode is HIPAA-compatible: no audio leaves the device, no third-party vendor processes the recording, and you choose whether any AI provider sees the text. Compliance itself is a property of the practice and its policies — HHS does not certify software products — so "HIPAA-compatible" is the right frame: it fits the data-handling policies that HIPAA-style solo-dictation workflows need. It does not, and cannot, substitute for your agency's own privacy and security review. The HIPAA rules themselves are at hhs.gov/hipaa.
Can I dictate case notes on my iPhone during a home visit?
Not today. MetaWhisp is a macOS desktop app. An iOS version is on the roadmap for 2026 but is not shipped. For live field dictation today, you'd need a separate phone-based workflow and accept that the privacy story will be different from the local Mac setup. Many social workers handle this by jotting anchor facts in a paper notebook and dictating the full case note once they're back at their Mac.
How accurate is MetaWhisp on social work terminology?
We have one published first-party accuracy figure: 2.76% WER (~97% accuracy) on LibriSpeech test-clean. We have not benchmarked accuracy on social work vocabulary, mental-health terms, or county-specific form language. Expect some cleanup, especially on names, acronyms, and terms the underlying Whisper model hasn't seen often. The "Correct" and "Rewrite" processing modes help by letting your own AI key polish the transcript after the fact.
Do I need a paid plan to use MetaWhisp for case notes?
No. Free local mode is unlimited, has no time caps, and is the right tier for case notes if PHI containment is the priority. Pro ($30/year or $7.77/month) is only worth it if you want built-in cloud AI without managing your own API key, or if you need cloud transcription — and in either case, audio does leave your Mac. For most social work workflows, Free plus your own OpenAI key is the better choice. See the pricing page for the full breakdown.
Can I export case notes into my agency's EHR?
MetaWhisp copies the finished transcript to your clipboard. Pasting it into your EHR is a manual step, but that's the right place for a human review and any EHR-specific formatting. We don't currently auto-sync with Epic, Avatar, or other systems. If your agency wants a programmatic export, that's a custom integration we'd need to scope.
Will MetaWhisp work on my agency's managed Mac?
Probably, but you'll likely need IT to approve the install. MetaWhisp is a downloadable .app, not an MDM-rolled SaaS. If your agency restricts application installs, ask whether they permit unsigned or non-MDM tools; if they don't, your IT department becomes part of the decision. There's no special enterprise version of the app — the same build runs on personal and managed Macs.
What languages does MetaWhisp support?
Ninety-nine languages with auto-detect in the underlying Whisper large-v3-turbo model. If you serve clients whose primary language isn't English — a common situation in social work — MetaWhisp will transcribe their speech in the language they used and produce English text if you ask for translation. Translation is one of the AI processing modes and works on the free tier with your own API key.
About the author: Andrew Dyuzhov is the solo founder of MetaWhisp. He built the app on top of open-source Whisper and WhisperKit using AI coding tools, dictates daily in Russian and English, and runs the 7-app head-to-head transcription test cited in this article. He is not a lawyer, clinician, or social worker — if you have specific compliance questions about your agency, ask your supervisor.
