Anyone who wants to record a meeting and turn it into text today can choose between dozens of tools. Speech recognition has become good, and it gets better every year. Which is precisely why it is no longer the thing worth talking about.
What matters comes before and after. Before: the conditions under which a recording is made. After: what becomes of what was said. Both decide whether you end up with usable minutes or a long text file nobody reads.
Swiss German is a real hurdle, but not the biggest one
That a system reliably understands Bernese, Basel or Valais German cannot be taken for granted. It is work, and most general-purpose tools fail at it because their models are trained on standard German and treat dialect as noise.
This hurdle can be cleared. It is a question of training data, evaluation and patience.
The harder question comes next: in a Swiss board meeting, rarely is only one language spoken. An agenda item starts in standard German because the documents are written that way. The discussion slips into dialect as soon as it gets concrete. The colleague from Romandy answers in French. The technical terms are in English anyway.
This code-switching within a single sentence is the norm, not the exception. A system that can only switch between languages, but not mid-sentence, produces nonsense at exactly those points. And they are often the points where decisions are made.
What happens in real meetings
The following situations sound like edge cases. In practice they occur regularly, and each one can render a recording useless.
The recording breaks off. A mobile device loses its network in the middle of the discussion, the battery runs out, a tab is closed by accident, or the person recording has to leave early. If the recording only lives in memory and is lost when it breaks off, the meeting is gone. That is why live recording in HeyEla saves continuously, every five seconds. A dropout then costs seconds at most, not the meeting.
The meeting runs for eight hours. A strategy workshop with a two-hour lunch break in the middle is not one meeting in technical terms but two that belong together. If you do not plan for breaks, you get either two unconnected recordings or one with two hours of silence.
Seven hours of audio are sitting there at the end. That is a different task from forty minutes. Processing time, cost and error-proneness do not rise linearly, and anyone who has only developed against short test meetings notices this at the first real workshop.
The device is badly placed. A microphone at the head of a long table hears the other half of the board less well. That can be compensated for technically, but only if you plan for it.
Where the recording happens is a question of its own
Boards no longer sit only in one room. The same organisation records in the meeting room, next time in Microsoft Teams, the time after that in Google Meet, and at the quarterly meeting half the group is in the room and the other half on the call.
Each of these constellations sounds different and needs its own route into the system. A room recording has reverb and background noise but clean speaker changes. A conference recording has clean audio but latency and overlaps. A hybrid meeting has both at once, and the people in the room are quieter than those on the call.
Supporting only one of these cases means, in practice, supporting none. That is why HeyEla covers all three: recording in the browser and through the device for the meeting in the room, a bot that joins Teams and Meet sessions directly, and both together for the hybrid round.
From what was said to the minutes
Even a perfect transcript is not yet minutes. The difference is not linguistic but legal: Art. 713 CO requires documentation of deliberations and resolutions, not a transcript of the conversation.
In practice the difference is even greater than it sounds. In a board meeting, a resolution rarely sounds like a resolution. It sounds like "all right, let us do it that way". Before it lie twenty minutes of discussion, after it three words of decision.
Whoever has only the spoken word cannot tell those three words from the twenty minutes. Whoever knows the agenda, has read the documents and knows what the board deferred last time, can.
That is exactly where the difference between a note-taking tool and a governance tool lies: not in the larger language model, but in the context it knows. We have described separately why generic note-taking tools fail on further counts in a board setting.
What stays with the board
Recording, transcription, structuring and the draft can be automated. Responsibility cannot.
Whether a formulation reflects the resolution correctly, whether a reservation belongs on the record, whether a conflict of interest must be documented: the board decides that. The AI proposes; review and approval are done by people. Under Art. 713 para. 3 CO, the chair and the person keeping the minutes sign at the end.
This is not a limitation but the right division of labour. It moves time to where it has an effect: away from typing, towards reviewing.
Conclusion
The question of whether an AI understands Swiss German is the wrong question today. It does understand it, with varying quality, and the gap is narrowing.
The right question is: what happens when the meeting does not go the way it does in the demo video? When the connection drops, when the language switches, when a meeting turns into a workshop. And what happens afterwards to what was recorded.
That is exactly where it is decided whether minutes emerge at the end.
Frequently asked questions
Can AI transcribe Swiss German? Yes. Quality differs markedly depending on dialect and recording situation. What matters is less whether a system understands dialect in principle than how it handles language switches within a meeting.
What happens if the connection drops during the meeting? With HeyEla the recording is saved continuously, every five seconds. A dropped connection, an empty battery or a closed tab therefore costs the last few seconds at most, not the meeting.
Does recording also work in Microsoft Teams or Google Meet? Yes. A bot joins the Teams or Meet session directly and records. Meetings in the meeting room you record through the browser or the device, and hybrid rounds with participants in the room and on the call work just as well.
What language are the minutes written in for a multilingual meeting? The minutes are produced in the minute-taking language defined by the organisation, regardless of which languages were spoken in the meeting. The Swiss Code of Obligations does not prescribe a language for the minutes.