The short answer
Dictating privileged material into a cloud transcription tool places client communications on a third party's servers. That is not automatically improper: bar guidance has generally allowed cloud tools where the lawyer takes reasonable care, but it converts confidentiality from something your office controls into a vendor-management obligation: vetting, contracts, retention, breach exposure. On-device transcription removes that third party from the picture entirely. This piece covers what to ask before adopting any dictation tool, and when the local option is simply the easier answer.
What actually happens to dictated audio in the cloud
Most dictation and transcription apps record your speech, upload it to the vendor's servers for recognition, and return text. For routine content that is a reasonable design. For privileged material it means:
- A new party holds client communications. The vendor's staff, infrastructure, and subprocessors are now in the chain of custody.
- Retention you have to manage by contract. Deleting your transcript does not necessarily delete the vendor's copies, and defaults vary.
- Terms that can permit training use. Some consumer-tier services reserve rights to use customer content to improve their models unless you opt out.
- Discovery and breach surface. Data held by a vendor adds a breach and disclosure surface that you, not your office, have to manage.
The duty in play
Lawyers owe clients competence and confidentiality, and ethics guidance has long framed technology choices as part of that duty: understand the tool, take reasonable measures to protect client information, and supervise the vendors you bring into the engagement. None of that forbids cloud tools. All of it is work: security questionnaires, data-processing terms, retention schedules, and an answer ready for the client who asks "who else heard this?" (This is general background, not legal advice for any specific engagement; your jurisdiction's rules and your client agreements govern.)
Five questions to ask any dictation vendor
- Is audio processed on my machine, or uploaded to your servers?
- If uploaded: how long are audio and transcripts retained, and can deletion be verified?
- Is customer content ever used to train models, and is the default off?
- Which subprocessors touch the data, and where is it stored?
- Will you sign terms consistent with my confidentiality obligations?
A tool that transcribes on-device lets you answer the first question with "neither audio nor text ever leaves my computer", which makes most of the rest moot for the transcription step.
The on-device alternative
Modern speech recognition runs well on an ordinary Windows PC, so the cloud round trip is no longer a technical necessity. With local transcription, privileged audio stays inside the security perimeter you already maintain for client files: there is no transcription vendor to vet because there is no transcription vendor. For how the technology works, see how local speech-to-text works; for the broader privacy picture, the voice transcription privacy guide.
Verify, don't take the brochure's word
"Local" is a testable claim: disconnect from the network and dictate. A genuinely on-device tool keeps working. MurmurTone documents every network call it makes (update checks, license validation, optional model downloads, the purchase flow) on its network transparency page; the dictation path makes none. The procedure is written up in how to check that a dictation app is actually offline.
Where MurmurTone fits for legal work
MurmurTone is Windows dictation built on local Whisper models: 100% offline recognition, custom vocabulary for party names and terms of art, text typed straight into your document system or email, $49/year with a 14-day trial (no credit card). Privileged audio never leaves the machine, so there is nothing to add to the vendor list. For the day-to-day workflow, see dictation for lawyers.
Run the test yourself. Download the trial, open a matter note, disconnect from the network, and dictate.