Board meetings, sales calls and project reviews produce plenty of decisions that later disappear into someone’s notebook. Self-hosted meeting transcription solves this without sending recordings to an external service: Whisper turns speech into text, and a local language model writes up the decisions and action items.
Below we explain how the setup works, how much hardware it really needs, how to handle participant notice and the GDPR, and how to roll it out without a months-long project.
The problem: meeting outcomes get lost or end up in the cloud
Manual notes are incomplete, and whoever writes them is only half in the conversation. Popular cloud meeting assistants do the job automatically, but the recording, transcript and summary all land on the provider’s servers. Board meetings involve financial figures and names, sales calls involve pricing, and a doctor’s appointment involves health data. Many organisations do not want to, or are not allowed to, move that content outside their own infrastructure. We covered this risk in more depth in our article on avoiding data leaks to public AI tools.
How self-hosted meeting transcription works
The whole process uses two models running on your GPU server and a simple application that accepts a recording and returns a finished set of notes.
Step 1: Whisper turns speech into text
Whisper is a speech recognition model released by OpenAI as open source under the MIT licence. It supports many languages, including Polish, German, French and Spanish. The whisper-large-v3-turbo variant is a faster version of the largest model, at the cost of a small drop in accuracy. It has fewer than one billion parameters, so compared with language models it is small and fits on a single GPU alongside one.
Step 2: a language model writes the notes
A raw transcript of a one-hour meeting runs to many pages that nobody will read. A second, locally hosted model turns it into what you actually need: a summary, a list of decisions, action items with an owner and deadline, and open questions. For this job, mid-size multilingual models such as Qwen3.8-27B (Apache 2.0) work well, and for meetings held only in Polish, small Polish models such as Bielik are an option.
Who said what
Whisper does not distinguish between speakers on its own. If the notes need to attribute statements to people, you add a third, small model for speaker diarisation (pyannote, for example). For online meetings it is often simpler to record each participant on a separate track.
If the notes should go straight into your CRM, task tracker or participants’ inboxes, you can automate that step with n8n running on the same server.
Notice and the GDPR: a recording is personal data
A voice recording and the content of what someone says can identify a person, so they are personal data. Your own server solves the problem of handing data to an external provider, but it does not remove your other obligations. In practice:
- Inform participants before recording that the meeting is recorded and transcribed, why, and who will see the notes. This includes external guests.
- Establish the legal basis with your data protection officer. Consent is only one option and not always the best one, for example between employer and employee.
- Set a retention period. A sensible practice is to delete the audio once the notes are approved and keep only the text.
- Restrict access to recordings and transcripts to the people who need them, and log that access.
In healthcare, a conversation with a patient contains special-category data and the requirements are stricter. We cover this separately in AI in healthcare clinics. For the GDPR and the AI Act with self-hosted models more broadly, see Self-hosted AI and EU compliance.
What hardware is enough
A single GPU is enough for a team. Transcription does not have to run in real time: recordings go into a queue and are processed one after another, so even a single card can handle many meetings a day. The card’s memory mainly determines how large a language model you can fit next to Whisper.
| Setup | GPU memory | Suitable for |
|---|---|---|
| Entry-level VPS with GPU | 12 GB | Testing, small teams, Whisper plus a small notes model |
| Mid-range VPS with GPU | 24 GB | A team or department, Whisper plus a 7–12B model |
| Larger VPS with GPU | 48 GB | Larger organisations, a bigger model and better summaries |
| Dual-GPU VPS | 2× 48 GB | Several departments, one card for transcription and one for the language model |
All of these are available as ready-made VPS packages with a dedicated GPU, listed on our DeepSeek page. Despite the name, they run any model, including Whisper. A quick rule-of-thumb check: an 11B model at 4-bit precision needs about 5.5 GB for its weights, so together with Whisper and headroom for context it fits comfortably on a 24 GB card. The web application and recording storage can sit on a standard VPS server, with the GPU used only for compute. To serve the language model to many users, vLLM is a better choice than Ollama; we explain the differences in Ollama vs vLLM vs SGLang vs NVIDIA NIM.
Rollout step by step
- Pick one meeting type, such as the weekly management meeting or sales calls, and define what good notes look like.
- Sort out the paperwork: privacy notice, legal basis, retention period and who has access.
- Run a pilot on a GPU VPS package: Whisper, a language model and a simple upload form.
- Check quality on real recordings made with your usual audio equipment, jargon and proper names. Refine the notes template.
- Connect the tools you already use: calendar, task tracker, CRM.
- Hand over operations: run updates, backups and monitoring yourself with root access, or let us handle them under our 24/7 administration.
Pitfalls
- Audio quality. A laptop microphone at the end of the table and several people talking at once are the main sources of errors. A decent conference microphone often helps more than a bigger model.
- Proper names and jargon. Surnames, product names and industry abbreviations get mangled. A vocabulary list passed to the model and a quick proofread before approval both help.
- Invented text. Whisper can produce sentences nobody said, usually during long stretches of silence. Trimming silence before transcription helps reduce this.
- Unreviewed notes. The model may assign a task to the wrong person or drop a caveat. The meeting chair should approve the notes.
- Recordings kept forever. An archive of recordings grows quickly and becomes a liability. Set up automatic deletion.
FAQ: frequently asked questions
Can Whisper run locally without internet access?
Yes. Once the model is downloaded, Whisper runs entirely on your server, so recordings never leave it. The same applies to the language model that writes the notes.
Is it legal to record meetings for AI transcription under the GDPR?
It can be, provided participants are informed before recording and you have a valid legal basis. Whether that is consent or another GDPR basis depends on the relationship and purpose, so agree it with your data protection officer.
What GPU do I need for Whisper and meeting summaries?
A single GPU is enough for a team. A 24 GB card is a comfortable starting point that fits both Whisper and a language model for the notes.
Does self-hosted transcription work in real time?
It can, but notes do not need it. Processing after the meeting is simpler and lets the same hardware handle more recordings.
Can Whisper tell who is speaking?
No. To attribute statements to people you need an additional diarisation model, or you record each participant on a separate track.
Want self-hosted meeting transcription that keeps recordings inside an EU data center? Start with one of our ready-made GPU VPS packages, and we will help you set up Whisper and the notes model. Contact us.