Practical guidance, not legal advice. Your duties come from your own jurisdiction's rules of professional conduct and your firm's policies.
Why the tool choice is a professional question
When you dictate into a cloud service, the audio goes to a third party that stores it, processes it, and in some cases uses it to train models. Whether that is acceptable is not a matter of taste — it engages confidentiality duties, and in some circumstances the question of whether disclosure to a vendor affects privilege. Firms handle this with vendor diligence and contractual terms, which works, but it is work, and it has to be redone for every tool and every renewal.
Offline software sidesteps the question: if the audio never leaves the machine, there is no third-party disclosure to assess.
What “offline” should actually mean
The word is used loosely. Insist on the specifics:
- Where does recognition run? On the device's own processor, or on a server the vendor operates? “Private cloud” is still a cloud.
- Does it work with the network off? The decisive test. Pull the cable, dictate, and see. Most tools need a one-time model download; after that, real offline software keeps working.
- What else does it send? Licensing, update checks and crash reporters are separate channels from the audio. Ask what each carries.
- Is there telemetry? Usage analytics that leaves the machine can carry more than a vendor intends — document titles, application names.
- Can the engine be audited? An open recognition model can be inspected by your IT team; a proprietary one has to be taken on trust.
The practical criteria, beyond privacy
A tool that is private but unpleasant will not get used, and an unused tool saves nobody any time.
- It must insert text where you are already working — the brief, the email, the case management system — not into a separate window you then copy from.
- Names and terms of art must survive. Party names, statutes and Latin phrases are where general-purpose recognition fails. A custom vocabulary matters more than a headline accuracy figure.
- Formatting should match the context. An attendance note and a message to a colleague do not want the same punctuation and tone.
- Cost should be predictable. Per-seat monthly billing for a typing aid adds up quietly across a firm.
Questions to put to any vendor in writing
| Question | Why it matters |
|---|---|
| Is audio or text ever transmitted off the device, and to whom? | Establishes whether a disclosure occurs at all |
| Do you retain recordings or transcripts? | Determines whether a third party holds client material |
| Do you train models on customer data? | Training corpora are difficult to unwind after the fact |
| Does the software require an account? | An account is another record linking you to your work |
| What is logged, locally and remotely? | Logs are discoverable and often overlooked |
| Can our IT team verify these answers independently? | A vendor confident in its architecture will publish it |
Where Altypist stands on each
Altypist is on-device dictation for macOS. Audio and text never leave your Mac; recognition runs on the Apple Neural Engine using the open-source Whisper model, so the engine itself is publicly auditable. There is no account, no telemetry and no analytics. The only network connections it makes are a one-time model download, licence activation — which sends a licence key and a one-way hash of the hardware identifier, and nothing else — and an update check. Text is inserted straight into the field where your cursor already is, a custom vocabulary handles party names and terms of art, and it is bought once rather than rented monthly.
The full architecture is published, in the form a compliance reviewer can work through.
Test it against the criteria above. 7-day trial, every feature enabled, no credit card and no account. Turn the network off during it — that is the check worth doing.