Three steps: text, who speaks and summary
First a speech model turns the audio into text. Then another separates the speakers, so the minutes say who said what. Finally a language model writes the summary: decisions, tasks and dates.
The first two steps are light. The third is the one that needs memory, because the model has to read the whole meeting at once.
Why do it in the office
Voice is personal data, and in a consultation or a case file it also carries health data or data protected by professional secrecy. Processing it on your own machine avoids sending it to a third party and makes clear where the audio is and who has access to it.
What hardware you need
For transcription, almost any modern computer will do. What makes the difference is the summary: a two-hour meeting is many pages of text, and the model needs memory to read them together.
| Case | Computer | Why | Price |
|---|---|---|---|
| One meeting room or office | Compact AI 32 | Transcribes and summarises meetings one at a time | €2,065 |
| Several offices or long meetings | Compact AI 64 | A larger summary model and several transcriptions queued | €2,940 |
| Many recordings a day | Compact AI 64 with RTX PRO 2000 | The graphics card transcribes hours of audio in minutes | Custom |
| Plus an assistant on your documents | Compact AI 128 | The same machine handles the minutes and searches the archive | €4,820 |
How it is set up
The compact stays in the office, connected to the network. Each person uploads their recording from the browser, or the room microphone sends it automatically. A few minutes later the text and the draft minutes are in their folder.
Nothing needs to be installed on the office computers. The machine has remote management over IP, so it is looked after at a distance.
What always has to be reviewed
Speech models are very accurate, but they slip on proper names, acronyms and figures, and get confused when two people speak at once. The minutes are a draft: someone who was in the meeting has to read them before signing.
A good microphone improves the result more than a larger model.
Related hardware and guides
Frequently asked questions
Does it work in several languages?
Yes. Open speech models transcribe the main European languages with good quality and recognise many others. In meetings where languages are mixed, the result needs a closer review.
Is a dedicated graphics card needed?
For one meeting room or office, no. With many hours of audio a day, a professional graphics card cuts the wait from hours to minutes.
Can it transcribe live?
Yes, with some delay. The most reliable approach is still to record the meeting and process it afterwards, because the model is more accurate when it hears the whole sentence.
Where is the audio stored?
On the machine itself or in the network folder you choose. Nothing leaves the office. You can set the audio to be deleted automatically after the number of days you decide.
What will you use it for?
Tell us which models you want to run, how many people will use it and where it will work. We will propose the compact, with or without a professional graphics card, and its price within 12 working hours.