Meeting transcription without the cloud: GDPR guide
In short. Transcribing a meeting without the cloud means the audio is turned into text on your own machine, without going through a third-party service. That reduces one risk but does not settle everything: you still need to inform the participants, set a retention period and proofread the result. This guide gives an 8-point check to run before you record.
Meeting transcription saves time, but a meeting transcript is also a sensitive file: a voice, names, sometimes figures or internal remarks. So the question is not only “which tool?” but “where does the audio go, who can read it and how long do we keep it?”. This guide draws on the Whisper README, three CNIL pages and the text of the GDPR, all read on 28 October 2026.
What are the three ways to transcribe a meeting?
| Assistant that joins the call | Phone app or online service | Local transcription | |
|---|---|---|---|
| Where the audio goes | to the assistant’s provider | to the service provider | stays on your machine |
| Do participants see it | often a banner or a “bot” participant | no, unless you tell them | only if you tell them |
| Contract | provider’s terms, to be read | provider’s terms, to be read | you define it (retention, access) |
| Cost shape | per-user subscription | free or subscription | hardware and setup, no per-minute charge |
| Weak point | the audio leaves the company | same, with less control | setup, quality to be tested, maintenance |
Here, local means a fully local workflow, not just a locally installed app. Check external APIs, cloud fallback, telemetry and automatic folder synchronisation. Also review backup destinations, software licences and any support-provider agreements. Hardware, setup and human review still cost money; local does not always mean free.
No option is “the right one” in itself. For a low-stakes meeting an online tool may be enough. For client interviews, patient files or exchanges with a lawyer, ask where the processing happens before you start, not after.
What do the CNIL and the GDPR say?
The three CNIL pages below cover phone-call recordings and remote work; they are not a complete rulebook for every meeting. Informing participants is not, by itself, a legal basis. Identify the appropriate legal basis and any additional duties with your data protection officer or lawyer before recording.
-
Inform people. For recorded phone calls the CNIL asks you to state at the start the purpose of the recording, who receives it, and the rights to object and to access. The same transparency is a good base for a meeting.
-
No permanent monitoring. In its remote-work Q&A the CNIL treats continuous video or audio monitoring of employees as incompatible with data protection, and notes that switching on a camera is personal-data processing. An announced meeting can still involve employee monitoring, depending on its purpose and use: you transcribe an announced meeting, not an employee’s working day.
-
The principles of GDPR Article 5. Data is collected for “specified, explicit and legitimate purposes”, limited to what is necessary, and kept in identifiable form “for no longer than is necessary”.
In practice: a written purpose (for example “to write the minutes”), prior information, a retention period set in advance. The rest (legal basis, external participants, voice recordings) needs an analysis specific to your situation.
One practical consequence: the CNIL page lists a right of access to recordings, so it is prudent to assume the same applies to the transcript. Keep the audio and the text somewhere you can find them, and apply the agreed retention schedule, including any justified legal retention duties. A transcript forgotten on a shared drive is the usual way a short-lived file becomes a long-lived one.
What should you check before pressing record?
This is the list we propose. It is not legal advice; it is a starting point to adapt.
| Point | Question to ask | Good practice |
|---|---|---|
| 1. Purpose | Why are we recording? | one written sentence, such as “minutes” |
| 2. Necessity | Would manual notes be enough? | if yes, do not record |
| 3. Information | Do participants know we record? | say so at the start and in the invitation |
| 4. External participants | Are there clients or third parties? | tell them and accept a refusal |
| 5. Place of processing | Where is the audio turned into text? | on a machine you control |
| 6. Access | Who can open the audio and the text? | a named list, not an open shared folder |
| 7. Retention | When is the audio deleted? | for example after the minutes are approved |
| 8. Proofreading | Who checks the text before it is shared? | one named person, checking names and figures |
If any answer is “we don’t know”, decide before you start recording.
Which model should you use for local transcription?
Whisper is a speech recognition model under the MIT licence. Its README lists tiny, base, small, medium, large and turbo. Approximate video-memory figures are 1 GB for tiny, 10 GB for large and 6 GB for turbo. These estimates describe the project’s implementation, not every Whisper-based application. Turbo is intended for transcription, not translation. Installation requires pip and ffmpeg; test quality on your own recordings.
The takeaway: the bigger the model, the more memory it needs and generally the more reliable the output, but the gain depends on your language, accents, noise and number of voices. We give no universal accuracy figure because there is none: you measure it on your own recordings. The same logic applies to text models run locally: our guide to a local LLM for small business works out file sizes from the quantization tables.
How do you test a transcription before real meetings?
A proposed procedure, to run on a fictional recording whose participants have all agreed:
- Record five minutes of realistic conversation with two or three voices, figures, proper names and one technical term from your trade.
- Transcribe it with a mid-size model, then with a larger one.
- Compare with a reference text you wrote yourself: note misspelt names, wrong figures and missing passages.
- Locate the audio, transcript and temporary files, and test deletion of the working copies. Document backup retention and how deleted files are handled after a restore; do not assume that all backup copies disappear immediately.
- If offline operation is required, install the model and dependencies first, then disconnect and repeat the test. Check that no cloud fallback is required.
This is a procedure to follow, not tests carried out for a client. The result depends on your hardware and your recordings.
What are the limits of local transcription?
-
It is not a compliance certificate. You remain responsible for the processing. Local transcription does not replace informing participants or your data protection officer’s advice.
-
Errors remain possible. Proper names, figures, rare words and overlapping voices are common weak spots. Proofread before sharing.
-
Speaker labels are not guaranteed. Knowing who said what needs an extra step, with its own errors.
-
You need hardware and maintenance. A recent laptop may be enough for short recordings; for hours of audio or several users, a dedicated graphics card is often useful.
-
A transcript is not minutes. Decisions and follow-up actions are written afterwards, with human validation.
How does Castel Studio work on transcription?
We transcribe audio and video recordings from Nice: readable text with timestamps, speakers told apart (without inventing identities) and uncertain passages flagged, checked by a person against the audio. The starting price is 90 € per batch: up to 60 minutes in 3 files, one main language and up to 2 people, with usable sound. Where the processing happens, storage, access and retention are agreed before any files are sent; if processing must run on your own computer, say so in the request, and the setup is then quoted separately. For regular use, the local AI page also offers an assistant on your own documents from 790 €. Before choosing between an online offer and a local one, read our guide to private AI for business, which sets out the questions to put to a provider, and our article on PDF invoices to Excel for another example of local processing.
Describe the type of recordings and their length in a few lines, without sending sensitive files through the form. We answer in writing within 48 hours, including to tell you that an online tool is enough in your case.
Local transcription can avoid a per-minute charge, but not every expense: hardware, electricity, maintenance and proofreading time still have to be planned.
Questions
Can I record and transcribe a meeting without telling the participants?
Assume you must inform participants before you start: purpose, recipients, retention period and rights. The CNIL asks for this for recorded phone calls, and the same transparency logic applies to a meeting. Secret recording is a legal question: ask a lawyer or your data protection officer.
Is a local transcription GDPR compliant?
Not automatically. A verified, fully local workflow avoids sending audio to a third-party service, but purpose, information to people, retention, access and backups still need handling. Local reduces one risk; it does not replace a compliance analysis.
Which model should I use to transcribe a meeting locally?
Whisper, released under the MIT licence, is a common starting point. According to the project README its sizes run from tiny to large, needing roughly 1 to 10 GB of video memory. Quality varies with language, accent and sound quality, so test on a representative recording before choosing.
How long should I keep the audio and the transcript?
The GDPR asks you not to keep identifiable data longer than necessary. A prudent habit is to set the period in advance, for example deleting the audio once the minutes are approved, and to write it in the information notice. The exact period depends on your purpose.
Does a transcript replace meeting minutes?
No. A transcript is raw text that can contain errors in names, figures and who said what. Minutes pull out the decisions and actions. Always proofread before sharing, especially figures, dates and proper names.