Transcription in workshops: browser, local model or cloud, and what GDPR expects

Written by the IdeaHone Editorial Team · · 5 min read

Sculpted sound wave on a pink background

You want to capture what is said in the workshop so nobody has to type while people think. Then the data protection officer asks where the audio goes, and you do not know. This piece explains the three ways speech becomes text, what GDPR expects, and gives you a checklist to clear before anyone presses record.

Why the engine matters

The three engines differ exactly on the question your data protection officer and works council will ask: does the audio leave the room, and is it stored? Pick the engine first and the legal conversation gets much shorter.

Three ways speech becomes text

Browser speech recognition

Chrome and Edge have speech recognition built in. The browser sends the audio to the browser maker's servers, recognition happens there, and the text comes back. Fast, good quality, nothing to install. The price: the audio leaves the device and the room.

A local model in the browser

A speech model such as Whisper can run on the laptop itself, via WebGPU or WebAssembly. The model, some hundred megabytes, is downloaded once, and the audio never leaves the computer. Quality depends on the device, and old laptops fall behind. For confidential topics this is the engine to choose.

Cloud transcription APIs

Audio is uploaded to a transcription provider and text comes back. Best quality for many languages. Here the provider matters more than the technology: EU hosting, a data processing agreement, and a contract that rules out training on your data.

In all three cases the decisive question is not the engine but what happens to the raw audio. A tool that keeps only the text and discards the audio immediately is in a different category from one that stores recordings. Some tools are built that way (IdeaHone, for example, defaults to the local model for organisations that switch on EU-only processing, and stores text, not audio).

What GDPR asks for

None of this is legal advice, only the points your data protection officer will raise.

  • A legal basis. In an employment context, consent is weak because of the power imbalance: employees can hardly say no to the facilitator. Legitimate interest with full transparency is often the better route. Let the data protection officer decide.
  • Information before listening starts. Participants must know what is captured, what is stored, for how long, who can see it, and that they can ask for correction and deletion.
  • Voice profiles are biometric data. Attributing a sentence to a person from their voice is processing a special category of data. That needs explicit consent per person, freely given and revocable at any time. Text-only capture does not.
  • A defined retention period. Months, not forever. Write the number down and have the tool enforce it.
  • A data processing agreement with the tool provider, whichever engine you use.
  • The works council. In Germany the works council has a say on tools that could monitor performance or behaviour (BetrVG). A record of who said what can qualify. Involve them early.

What to say before you press record: "We capture what is said as text so we can build the results together. No audio is stored, only the words. If you do not want your name attached to anything you say, tell me and it stays anonymous."

Checklist before the session

  1. Pick the engine by sensitivity. Strategy, personnel or customer data: local model. Routine ideation with nothing confidential, like a 90-minute ideation session: browser or cloud, with the right contract.
  2. Inform participants in writing. What is captured, what is stored, for how long and who sees it, in the invitation. One paragraph is enough.
  3. Ask for consent only for voice profiles. Explicit, voluntary, revocable, per person. Everything else runs on the legal basis your data protection officer chose.
  4. Set retention. Agree a number of months and configure it in the tool.
  5. Sign the data processing agreement. With the tool provider, before the first session, not after.
  6. Inform the works council where it applies. In Germany, before the tool is used, with a short description of what it records.

How transcripts drift out of compliance

  • Treating a transcript like notes. Notes are one person's summary. A transcript records what people actually said, including the half-formed thought. Handle it with the care you would give a recording.
  • Switching engines without telling anyone. The team agreed on the local model and someone turns on cloud transcription for better quality. The information you gave is now false.
  • Voice attribution on by default. Keep it off until each person has said yes. Without names, a transcript still serves most workshops.

Before you choose a tool, go through the questions your data protection officer will ask.

Capture every idea while people talk

IdeaHone listens in your session and turns what people say into idea cards, questions and decisions, live on a shared board.

Request a demo

Yes, with a legal basis, information to participants beforehand, a defined retention period and a data processing agreement with the tool provider. Check the details with your data protection officer.

Not necessarily for text capture; in companies, legitimate interest with transparency is often the stronger basis because consent from employees is rarely free. You do need explicit consent per person for voice attribution, because voice profiles are biometric data.

Is a transcript a recording?

A transcript is text, not audio, but still personal data and a verbatim record of what people said. If the tool also stores the audio, you have a recording and the stricter rules apply.

All resources