Enterprise Compliance GDPR HIPAA

Enterprise Offline Voice-to-Text: Why Regulated Industries Need On-Device Speech Recognition

Regulated teams need on-device speech recognition because a voice recording is sensitive data, and every cloud dictation service adds a third party that receives it. If the audio is transcribed on the laptop and never sent, most of the compliance questions disappear. This guide explains why, sector by sector, and what an on-device rollout looks like. For reference: AirTypes Plain is free and transcribes on the endpoint with 0 bytes of audio uploaded; Polished, the optional cloud cleanup at 120 minutes for every dollar, is a per-user choice many organisations will keep away from sensitive work.

This article explains common compliance reasoning. It is not legal advice; your DPO, privacy counsel and security team decide what applies to your organisation.

Why on-device dictation is now an enterprise requirement

Three things changed. First, open-weight speech models, OpenAI's Whisper family in particular, became accurate enough on ordinary laptops that the historical reason for cloud dictation (better models, more compute) mostly went away for everyday English dictation. Second, scrutiny of cross-border data transfers kept rising: Schrems II and the EU-US Data Privacy Framework, the EU AI Act's rules on biometric categorisation, the Digital Operational Resilience Act (DORA) in financial services, and NIS2. Third, board-level interest in generative AI made legal and security teams look hard at any tool that sends raw user content to a third party, voice included.

The result is a common preference inside enterprise IT and security teams: where a tool can run locally without losing what users need, it should. Voice-to-text is one of the clearest cases, because the local option now covers the everyday job well: typing what someone says into the app they are already using.

Voice as a regulated data class

GDPR: voice is personal data, and can be biometric

Under the EU General Data Protection Regulation, recorded voice is personal data when it can be linked to an identified or identifiable person. When voice is processed to uniquely identify someone, it is biometric data in a special category under Article 9, with a higher bar for a lawful basis. Many data protection officers treat raw recordings cautiously even for transcription, because the audio itself carries identity.

Practical impact: if a European team uses a cloud dictation service hosted elsewhere, the organisation typically needs a transfer mechanism, a DPA, a transfer impact assessment, a view on how long the vendor keeps the audio, and possibly a DPIA. Most of that falls away when the audio is transcribed on the laptop and never sent.

HIPAA: clinical voice notes routinely contain PHI

In US healthcare, audio captured by clinicians, scribes and researchers usually contains Protected Health Information. A cloud speech vendor that handles PHI is a Business Associate, which means a Business Associate Agreement and the usual administrative, physical and technical safeguards. On-device transcription keeps PHI on the speech path inside the organisation's own endpoint.

DORA, MiFID II and SOX: banking and finance

Financial firms work under several overlapping regimes. DORA expects them to manage ICT third-party risk. MiFID II sets record-keeping rules for certain communications. SOX governs internal control over financial reporting. None of these bans cloud dictation outright, but each adds work to a vendor relationship that handles raw voice. A tool whose speech path never leaves the endpoint is a shorter conversation.

Export control and IP rules

Defence, aerospace, semiconductor and some energy firms work under export-control regimes such as ITAR, EAR and EU dual-use rules. Depending on jurisdiction and content, speaking technical data into a third-party speech service can raise export questions. On-device transcription keeps that data where it already is.

Automotive and manufacturing: where engineering speech is competitive intelligence

Vehicle makers and their suppliers run large engineering organisations whose everyday writing includes prototype test results, supplier negotiations, regulatory dossiers and pre-launch product detail. Engineers often want voice-to-text, because narrating a requirement or a test observation is quicker than typing it. When the only options send audio to a cloud service, the answer from IT is often no.

An on-device option changes that conversation. The audio stays on the engineer's laptop, the text goes into the same requirements tool or document as before, and nobody has to argue the case for a new data processor. The same logic applies to battery, semiconductor and industrial suppliers whose roadmaps are themselves sensitive.

Banking and finance: MNPI, internal memos and "no audio off the endpoint"

Banks and asset managers typically have explicit policies that keep Material Non-Public Information off consumer cloud services. Voice notes about a pending deal, a draft research note, a credit memo or a pre-earnings preview all fall into that category. The dictation tool that passes an information-security review is one whose data flow is simple to state: audio is captured, transcribed locally and discarded, and the text stays in the same controlled application as before.

That describes AirTypes in Plain mode. It does not describe Polished, which sends the recording to AirTypes; a bank would reasonably keep Polished away from anything covered by an MNPI policy.

Healthcare and pharma: PHI, trial data, lab notes

Hospitals, academic medical centres and pharmaceutical companies have established processes for PHI and trial data. Cloud dictation can be approved under a BAA with a specific vendor, but every approval takes time, and every new project or site can mean another one. On-device dictation keeps the speech path inside the endpoint, which narrows what has to be reviewed.

One caution: Plain is English-only, general dictation. It is not a medical transcription product and has no specialist medical vocabulary. Clinicians should proofread terms, doses and names, as they would with any dictation tool.

Public sector and defence

Government and defence organisations handle information at classifications where cloud dictation is often not permitted at all. On-device transcription is the only shape that fits, and the requirement is usually air-gap friendly: the speech model must run without a network connection. Plain does that once the app is set up and its model is downloaded. Whether an app may be installed on a given network at all is, of course, a decision for the organisation's own accreditation process.

Reference architecture for an offline dictation rollout

The shape that is easiest to defend for on-device dictation looks like this:

  1. Endpoint-only audio. The microphone is captured locally and the audio never leaves the device. No waveform is sent for "model improvement".
  2. A local Whisper-class model. The model is downloaded once and runs on the endpoint's CPU. No GPU is needed.
  3. Text injection at the cursor. The text is typed into whatever app has focus, using the operating system's keyboard input. The dictation tool needs no integration with the receiving app.
  4. Any cloud step is separate and visible. If the tool offers cloud cleanup, it is a distinct mode the user switches on, not something that happens silently, and users can work without it.
  5. No network needed on the dictation path. Sign-in may talk to the vendor's servers, but transcription works with the network off.

This pattern is easy to explain in a security review: on the local path there is no third-party processor of voice data, no cross-border transfer of voice data and no retention of voice data by a vendor.

How AirTypes fits the pattern

AirTypes is a desktop dictation app for macOS and Linux. The user holds Ctrl + Shift + Space (customizable, hold-to-talk or toggle), speaks, lets go, and the text is typed at the cursor in any desktop app through native keyboard input. There is no plugin and no Office or Slack integration, so the vendor of the receiving app is not part of the review either.

Plain: the enterprise case

  • OpenAI's Whisper runs offline on the endpoint's CPU. 0 bytes of audio are uploaded. It works in airplane mode.
  • English only, with two model tiers: Swift (Whisper tiny, about 78 MB) and Crisp (Whisper base, about 148 MB), the one we recommend. Both run on a 4 GB machine with no GPU.
  • Light cleanup runs on the machine too: it removes a leading um or uh, capitalizes the first letter and adds a period.
  • Free, with no time limit. Users create a free account at first launch, with an email and password or Google.

Polished: a per-user choice to think about

Polished is the optional cloud mode. Polished sends the recording to AirTypes to transcribe and clean it up: a large speech model on our servers transcribes it, then an AI removes filler words and false starts and fixes grammar and punctuation. It works in the language the user speaks and runs on prepaid credit ($5 buys 600 minutes, $10 buys 1,200).

For a regulated team, the honest position is this: Polished is not part of the on-device argument. It is a choice each user makes on the app's Home screen, and it only runs while that user's account has credit. Every account gets its first minute of Polished free, once; after that it needs prepaid credit. An organisation may well prefer to avoid it for sensitive work. If so, the practical step is simple: tell users to stay in Plain and not to add credit. Once the free minute is used and there is no credit, Polished does not run and Plain keeps working. The same holds for Dictate in the browser on airtypes.com, which is Polished only: it sends the recording to AirTypes and spends the same credit.

What AirTypes receives

Always: the account email and usage counts. In Plain, nothing else. In Polished, also the recording. That is the full list, and it is the sentence to put in your risk register. For a broader look at how dictation tools handle data, see is voice dictation private?

Platforms

macOS (Apple Silicon and Intel) and Linux (.deb, .rpm and AppImage) are available today. The Windows desktop app is in development; teams standardised on Windows can join the waitlist. Dictate in the browser works on Windows today, but it is Polished only, so it is not an on-device option. Platform setup is covered in the guides for macOS and Linux.

A short risk-review checklist for procurement

If you are evaluating any voice dictation tool for regulated work, these are the questions that usually come up in security and procurement reviews:

  • Where is audio captured? (Should be: on the endpoint.)
  • Where is audio transcribed? (Should be: on the endpoint, for sensitive work.)
  • Is any audio sent anywhere, for "model improvement" or otherwise? (Should be: no, on the local path.)
  • Does transcription need an internet connection? (Should be: no.)
  • Does it work with the network switched off? (Should be: yes, once set up.)
  • If the tool has a cloud mode, is it a separate, visible choice, and can users work entirely without it? (Should be: yes.)
  • Exactly what does the vendor receive in each mode? (Should be: a short, written list.)
  • Is text inserted through the operating system or through per-app integrations? (OS-level avoids a review for each receiving app.)
  • Does dictation keep working during a network outage? (Should be: yes, on the local path.)

For AirTypes, the answers for Plain are: endpoint, endpoint, no, no, yes, yes (Polished is a separate mode that needs credit), "account email and usage counts", OS-level, yes.

FAQ

Is voice considered personal data under GDPR?

Yes. Recorded voice that can be linked to an identified or identifiable person is personal data. Voice processed to uniquely identify someone is biometric special-category data under Article 9, and many data protection officers treat raw recordings as high-risk because the audio carries identity. Ask your own DPO how your organisation classifies it.

Does on-device transcription remove GDPR obligations entirely?

No. The resulting text is still personal data where it is about people, and it must be handled under your normal data policy. What on-device transcription removes is the transfer of the audio to a third party, the speech processor relationship, and the retention question on the audio path.

Is AirTypes air-gap compatible?

Plain mode transcribes with no network: Whisper runs on the endpoint CPU, 0 bytes of audio are uploaded, and it works in airplane mode. The app needs a connection to install, sign in and download the voice model, and it checks the account when it starts, so open it online before you go offline. Polished mode always needs the internet, because it sends the recording to AirTypes.

What does AirTypes receive from a user's machine?

Always the account email and usage counts. In Plain mode, nothing else. In Polished mode, also the recording, because Polished sends the recording to AirTypes to transcribe and clean it up.

Can we stop staff from using the cloud mode?

Polished is a per-user choice made on the app's Home screen, and it only runs while that user's account has credit. Every account gets its first minute of Polished free, once; after that it needs prepaid credit. If your policy rules out cloud processing for sensitive work, tell users to stay in Plain and not to add credit. Once the free minute is used and there is no credit, Polished does not run and Plain keeps working.

Related: how offline speech recognition works, and the best offline speech recognition apps compared.

Get AirTypes, free

Plain is free and transcribes on the endpoint, with 0 bytes of audio uploaded. Polished is optional, per user, and never needed. macOS and Linux available now; the Windows app is in development.

Download for macOS and Linux