Meeting minutes with a local AI: from the recording to approved minutes, without the cloud

How AI meeting minutes work on your own server

AI meeting minutes on your own server work like this: the meeting is recorded, a speech recognition model turns the audio into text, a second model marks who spoke when, and a language model writes a draft of the minutes with decisions, tasks and open points. The minute-taker checks the draft against the transcript and approves it. All of it runs on a machine in your building or in a data centre you choose, so recordings of your management meetings, customer calls and project reviews go to no outside service.

We wrote this for the office manager, the executive assistant or the managing director who spends hours after every meeting on minutes and does not want to upload recordings to a cloud tool. The workflow comes first, then the questions of data protection that come with any recording, then the server and what it needs in the building. Our guides to the speech-to-text server for Whisper and Parakeet and to speaker diarisation on-premise cover the models in depth for the IT team; here the subject is the office task from start to finish.

Four terms come up throughout. A transcript is the written text of everything said. Speaker labels, which engineers call diarisation, mark the parts of the transcript as speaker 1, speaker 2 and so on. A language model is the AI that reads the transcript and writes the minutes. A GPU, or graphics card, is the part of the server that runs these models.

The six steps from recording to approved minutes

Our summary of the workflow; models as described in their Hugging Face cards, read 25 October 2026.

Each step works only with what the previous one produced: a wrong speaker label assigns a task to the wrong person. That is why the check in step 5 stays with a person.

Recording: room microphone or video call

The speech models in this guide read audio at 16 kHz on a single channel: the Parakeet model card names 16 kHz mono in WAV or FLAC, and NVIDIA's diarisation model the same sample rate. Any ordinary conference microphone delivers that. What changes the result is the room.

A meeting room with one table microphone records every voice on one channel, from different distances, with chairs, coffee cups and people talking over each other. This is the hard case. The Parakeet TDT 0.6B v3 model card shows how hard: on the AMI corpus, recorded meetings, its word error rate is 11.31 per cent, against 1.93 per cent on LibriSpeech test-clean, which is read speech. In plain words, about one word in nine is wrong on meeting audio, against one in fifty on a clean recording. The model card's noise test points the same way: the average error rate across its test sets rises from 6.34 per cent on clean audio to 11.66 per cent with noise as loud as the speech.

A video call usually gives cleaner audio, because each participant speaks into a headset or laptop microphone close to the mouth. Many conferencing systems can export the recording as a file. Where a system records each participant on a separate audio track, the tracks already say who spoke, and the speaker labelling step can be left out, as with separate channels in a contact centre; our guide to contact centre AI on-premise describes that case. Check the conferencing system's own transcript and summary features: where they run on the provider's servers, the recording leaves the company there, whatever you do with the file afterwards.

Three practical rules improve the transcript before any model runs. Put the microphone in the middle of the table, not next to the projector. Ask people to say their name before they speak for the first time in a large meeting. And keep side conversations out of the recording, because overlapping speech is where both recognition and speaker labels fail most.

Transcribing meetings locally in German and other languages

Two open models cover most European offices in October 2026. OpenAI's Whisper large-v3 lists 99 languages on its Hugging Face model card, which is tagged Apache 2.0, and OpenAI's repository gives about 10 GB of GPU memory for the large models. NVIDIA's Parakeet TDT 0.6B v3 transcribes 25 European languages, among them German, French, Italian, Polish, Czech and Romanian, detects the language by itself, adds punctuation and timestamps per word, and is published under CC BY 4.0. Its card states that it handles audio "up to 24 minutes long with full attention (on A100 80GB) or up to 3 hours with local attention" in one pass.

Open ASR Leaderboard paper (arXiv 2510.06961v3, 10 December 2025), German word error rates from its multilingual table, RTFx on its English test sets on one A100-SXM4-80GB; languages from the Hugging Face cards; all read 25 October 2026.

The German figures in that table come from public test sets, not from meeting rooms, so expect higher error rates on your own recordings, as the AMI figure above shows. Whisper's model card adds two warnings worth knowing. It says the models "perform unevenly across languages" and on different accents and dialects, and that "the predictions may include texts that are not actually spoken in the audio input (i.e. hallucination)". A minute-taker who reads the transcript catches such passages; a summary written from the transcript alone can carry them into the minutes.

Speed is not the issue. RTFx is the number of seconds of audio a model transcribes per second of computing. At Whisper large-v3's 145.5, a one-hour meeting takes 3,600 ÷ 145.5 = about 25 seconds of GPU time; at Parakeet's 3,333, about one second, by our arithmetic on the paper's A100 and English test sets. Even 20 one-hour meetings a day come to about eight minutes with Whisper. Your card and language change these figures, not their order of magnitude.

Who said what: speaker labels and names

Minutes need to know who agreed to what. A diarisation model marks each passage with an anonymous label and a start and end time. NVIDIA's Nemotron 3 Diarization, published on 23 September 2026 under the OpenMDW License 1.1, "handles up to eight speakers", and its card states it is "ready for commercial or non-commercial use". On AMI meetings recorded with a single distant microphone, its card reports a diarisation error rate of 11.14 per cent for the offline profile. For meetings with more than eight voices, pyannote community-1, under CC BY 4.0, lets you set a minimum and maximum number of speakers, as our diarisation guide describes.

The labels are generic. Turning "speaker 2" into a name can happen in two ways. The minute-taker assigns names by hand from the list of attendees, which takes a minute per meeting. Or the system compares voices with stored voice samples of known people. The second way produces data that the GDPR may treat as biometric: Article 4(14) defines biometric data as personal data "resulting from specific technical processing" of a person's characteristics that allow their unique identification, and Article 9(1) prohibits processing "biometric data for the purpose of uniquely identifying" a person unless an exception applies. For office minutes the manual way is simpler, and your data protection officer decides whether anything else is needed.

From transcript to minutes: what the language model does

The language model reads the transcript with its speaker labels and writes the draft along a fixed template: date, participants, agenda items, decisions, tasks with owner and due date, open questions. A good template asks the model to give the timestamp of the passage behind every decision and task, so the minute-taker can jump to the place in the transcript and check it.

A one-hour meeting fits comfortably. At an example value of 8,000 spoken words and the 1.76 tokens per German word measured in a preprint cited in our guide to tokens per word in European languages, the transcript comes to about 14,100 tokens, the units in which models count text. gpt-oss-120b, OpenAI's open model under Apache 2.0, accepts 131,072 tokens according to its configuration file, so a whole day of meetings would fit, and a session of 32,768 tokens holds one hour of transcript, the template and the draft.

Writing the draft is quick: gpt-oss-120b reads about 5 GB of weights per generated token, so one RTX PRO 6000 Server Edition with 1,597 GB/s allows about 320 tokens a second for one request, an upper bound by our estimate. The minute-taker's check takes longer than the machine.

Two rules keep the minutes reliable: decisions are written as the meeting stated them, and anything the model cannot place, such as a task without an owner, is listed as open instead of being filled in.

We build the servers for this task and set up the models with our Private AI/ML service; send us the number of meetings a week, the languages and how they are recorded, and our team replies with a configuration.

Recording meetings and data protection: informing participants

A recording of a meeting in which people can be identified holds personal data. Running the models on your own server removes the outside provider from the picture, as our article on local AI and GDPR explains, but the duties of the company stay.

Inform before you record. Article 13(1) of the GDPR requires the controller to give information "at the time when personal data are obtained", including "the purposes of the processing for which the personal data are intended as well as the legal basis for the processing" and, under Article 13(2)(a), "the period for which the personal data will be stored". For a recurring meeting this can be a short notice in the invitation and a sentence at the start of the meeting; guests from other companies are informed too.

Choose the legal basis. Article 6(1) lists six bases, among them consent and the legitimate interests of the controller. For employees, consent is a weak basis: the European Data Protection Board's Guidelines 05/2020 on consent state that, because of the dependency between employer and employee, it is "unlikely that an employee would be able to respond freely". Which basis applies is for your data protection officer to decide. National law can add rules on recording spoken words and on employee data, and where there is a works council it usually has a say, as our guide to the works council and an internal AI assistant describes.

No emotion analysis. Some meeting tools score the mood of participants from their voices. Article 5(1)(f) of the AI Act prohibits "the use of AI systems to infer emotions of a natural person in the areas of workplace and education institutions", with exceptions for medical or safety reasons, and the prohibitions have applied since 2 February 2025. A minutes system on your own server leaves such features out.

Assess the risk. Article 35(1) asks for a data protection impact assessment where processing "is likely to result in a high risk"; whether regular recording of internal meetings reaches that level is your data protection officer's call. Our guide to a DPIA for an internal LLM describes the document. This section is general information as of October 2026, not legal advice.

Storage and deletion

Article 5(1)(e) of the GDPR requires that personal data are "kept in a form which permits identification of data subjects for no longer than is necessary". For minutes that usually means three different periods for three different things.

Example retention rules for illustration; the periods for your company are set with your data protection officer under Article 5(1)(e) of the GDPR (EUR-Lex consolidated text, read 25 October 2026).

The deletion has to reach every copy. Under Article 17(1) a person can ask for "the erasure of personal data", and on a minutes server the data sits in the audio store, the transcript store, the search index if transcripts are searchable, the chat history of whoever asked the assistant about the meeting, and the backups. Set up the deletion as an automatic job when the system is built, so that nobody has to remember it. Keeping recordings out of the company-wide document search is the simplest way to keep the list short.

The server, how many meetings, what it needs in the building

Minutes are rarely the only use: the sensible set-up is the company AI server that also runs the chat assistant, document search and translation, as our guide to an AI server for a company of 10 to 100 people describes; four cards give further departments a card each.

Reference configurations on our AI servers page; sessions by our users-per-card rule with gpt-oss-120b (60.8 GiB of weights, 1.125 GiB of 16-bit cache per 32,768-token session); power by our estimate (cards, processors, plus 15 per cent for the rest of the system), to be confirmed by a measurement on the delivered system.

Two cards in a server. The entry server on our AI servers page has two NVIDIA RTX PRO 6000 Blackwell Server Edition cards with 96 GB each, two AMD EPYC processors, 384 GB of DDR5 ECC memory and two NVMe drives. By our users-per-card rule, which takes 90 per cent of the 95.6 GiB the driver reports, less 3 GiB per card and less the model's weights, gpt-oss-120b leaves 22.2 GiB on the first card for 19 sessions of 32K tokens at the same moment; a minutes job takes one of them for a short while. MIG, NVIDIA's way of splitting a card into isolated parts, divides the second card into four instances of 24 GB: one for Whisper or Parakeet, one for the diarisation model, whose file is 397 MB, one for the search models and one for tests. The server draws about 1.8 kW at the wall by our estimate and fits one 230 V 16 A feed, which carries 2.94 kW at 80 per cent. It belongs in a server room or a rack with cooling.

Two cards in a workstation. A company without a server room can use a workstation with two NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition cards, 300 W each, on an ordinary 230 V socket at about 1.1 kW by our estimate, best placed in a side room rather than an open office. The same split of tasks applies. Without a room of your own, our Private AI/ML service runs the platform on dedicated hardware in a Tier-3 data centre in Lithuania, inside the EU.

Notes for the IT team

1. Pipeline. Recognition, diarisation and the language model run as separate services; for recorded meetings use the diariser's offline profile, and measure GPU time per hour of audio with diarisation switched on, since our diarisation guide shows how much a combined pipeline can slow down.
2. Network. Keep the model engines on an internal network behind an LLM gateway with keys and a query log; vLLM's --api-key protects only some paths (vLLM 0.32.0 security guide).
3. No outbound calls. Check that no component sends audio to a hosted service; the pyannote card, for example, also offers a model that runs on its provider's servers.
4. Deletion jobs. Implement the retention rules as scheduled jobs that cover the stores, the index and the backup expiry.

What we supply

You can buy this system from us: our team sizes the configuration to your workload and budget and replies within one business day. We build AI servers with two or four RTX PRO 6000 Server Edition cards and workstations with two RTX PRO 6000 Max-Q to order, assembled and tested before delivery, delivered anywhere in the EU on one EU contract and invoice, with manufacturer warranty on every component handled through us; quantities and lead time are confirmed in writing with the quote. Our Private AI/ML service, with our engineering partner Vixen.UNO, sets up speech recognition, speaker labels, the minutes template and the deletion jobs; describe how your meetings are recorded today, and we propose a pilot with a few of your own recordings.

FAQ