AI transcription on our own servers: why we do it and what it means

Every AI feature in Coviqo uses speech recognition, translation and a language model that run on infrastructure we operate. Our servers are in Germany; the operator is based in Luxembourg. This covers:

• Speech recognition for live captions, meeting transcripts and transcripts of voice messages.
• Translation of captions and of meeting summaries into each reader’s language.
• The language model behind meeting summaries, suggested tasks and decisions, “Ask the space” and the catch-up paragraph in “What did I miss?”.

Audio, transcripts and other content are not sent to external AI providers, and we do not use them to train AI models. Our privacy policy says the same.

Meetings are where people say things they would not write down: numbers before they are final, staff matters, half-formed ideas. A transcript turns all of that into searchable text. Who processes that text matters.

If an AI feature calls an external service, a further company processes the content – often under its own terms, sometimes outside the EU. For a team or a club that means one more contract to check, one more entry in the list of processors and one more question from members that is hard to answer.

Running the models ourselves removes that extra party. The answer to “where does my meeting go?” stays short: to Coviqo’s servers, and nowhere else for the AI part.

Running AI on our own servers is only half of it. The other half is that people decide when it is used.

• Several switches must be on. AI features depend on the platform, the plan (the Free plan has no AI transcription or summaries), the space settings and the individual meeting. Space admins decide the defaults under Space settings › AI & transcription.
• Everyone can opt out. With Don’t transcribe my voice (Settings › AI & captions) your microphone is left out of captions, transcripts and summaries in every meeting. Your voice messages are not transcribed either.
• Everyone can see it. While transcription runs, a Transcribing badge is visible to all participants. Phone participants hear an announcement.
• Recording mutes first. When a recording starts, everyone except the person who started it is muted by default and sees a notice. Switching the microphone back on is a conscious choice.
• AI suggests, people decide. Summaries, tasks and decisions are drafts that are labelled as AI-generated. Suggestions are only published after a person has reviewed them.
• Some content is always excluded. Direct messages are never part of AI summaries, and end-to-end encrypted rooms have no AI features at all.
• You can check what happened. The privacy cockpit lists which AI feature processed content concerning you – without the content itself.

Transcripts are kept for 90 days by default in new spaces; admins can choose a different period.

Self-hosting AI is not free of compromises. We would rather name them than have you discover them.

• Capacity is finite. Speech recognition needs graphics cards, and we cannot add them in seconds. That is one reason why transcription minutes are a plan limit. If the service is briefly unavailable, the badge reads Transcription waiting, and transcription resumes on its own when the service is back.
• Models that fit our hardware. We choose models that run well on our servers. Summaries are useful drafts, but they can miss nuance or contain errors. Check important statements against the transcript or the recording.
• Languages are not equal. Recognition works in many languages, but quality depends on the language, accents, audio quality and people talking over each other. Setting the spoken language of a meeting instead of automatic detection often helps.
• Not everywhere. In end-to-end encrypted rooms the server cannot read the content, so there is no transcription or summary there. That is the point of those rooms.

• Set the spoken language when you schedule the meeting.
• Use a headset or a decent microphone, and let one person speak at a time.
• Write the agenda and key points into the meeting notes – they feed into the summary together with the transcript and the chat.
• Review suggested tasks and decisions before you publish them.
• Tell guests at the start that transcription is on, and point them to the opt-out.

All AI features are described on AI in Coviqo, and step by step in the guide chapter on meetings.