Every company knows the pattern: an hour of video call, three pages of scattered notes, and two weeks later someone asks what was actually decided on agenda point three. Cloud transcription tools solve the technical problem, but every word of every client call, board discussion, and HR interview travels to third-party servers. Under GDPR, that is a real compliance liability.
There is a better path: Whisper, OpenAI's open-source speech recognition model, runs entirely on your own hardware, no API key, no data transfer, no upstream vendor dependency.
What Whisper is
OpenAI released Whisper in September 2022 as an open-weights model trained on 680,000 hours of multilingual audio. The current flagship, Whisper large-v3, achieves a word error rate below 3 percent on clean English audio according to community benchmarks, comparable with commercial services like Google Speech-to-Text and Azure Cognitive Services. Practitioners report similar accuracy across German, Spanish, French, and other European languages when audio quality is reasonable.
Whisper turbo, an optimized variant, runs 6-8× faster than large-v3 according to community measurements, with minimal accuracy loss under real-world conditions. For meeting transcription, where completeness matters more than real-time processing, turbo is the practical default for most teams.
Two dominant runtime options
faster-whisper wraps the CTranslate2 engine for roughly 4× throughput over the original Whisper implementation on NVIDIA GPUs via int8 quantization, as reported by practitioners across the community. The built-in Silero VAD integration automatically skips silence and background noise, essential for office recordings with keyboard sounds, pauses, and overlapping voices.
whisper.cpp is CPU-optimized and works without any GPU. On Apple Silicon (Mac Studio M3 Ultra, MacBook Pro M3 Max), community users report transcription at 10-15× real-time: a 60-minute recording transcribed in under 6 minutes, on CPU alone.
The choice is straightforward:
- NVIDIA GPU environment → faster-whisper
- Apple Silicon or CPU-only Linux → whisper.cpp
Speaker diarization: who said what
Raw Whisper output is one undifferentiated text block. To attribute quotes to individuals, which every useful meeting record requires, you need speaker diarization alongside Whisper.
pyannote.audio is the dominant open-source option. It segments audio by speaker before Whisper transcribes each segment, producing a structured transcript with "[Speaker A]: ..." attribution rather than a wall of text. Practitioners report that the combination of pyannote + faster-whisper runs cleanly on a Linux server with an NVIDIA RTX 4090; Apple Silicon setups work but require more configuration effort.
A production-ready stack for an SMB
A meeting transcription pipeline for a 10-50-person team needs only a handful of components:
- Hardware: Mac Studio M3 Ultra (192 GB unified memory) or a Linux server with an NVIDIA GPU. A well-configured server with 32 GB RAM handles smaller teams on CPU.
- Model: Whisper turbo (multilingual) for speed; large-v3 for maximum precision on accented speech.
- Runtime: faster-whisper or whisper.cpp as appropriate.
- Post-processing: Ollama with a local LLM, Llama 3.3 70B or Qwen2.5 32B, to extract structured output.
- Integration: Structured minutes in Markdown or JSON, delivered via API to your existing tools.
The entire data flow stays on your hardware. Nothing leaves your network.
From raw transcript to structured minutes
Whisper produces text. A meeting record needs more: agenda structure, decisions, action items with owners and deadlines. A local LLM closes that gap.
A minimal workflow:
- Whisper transcribes the recording → timestamped raw text block.
- A local LLM (Llama 3.3 70B via Ollama) processes it: "Extract agenda items, decisions, and open tasks with responsible parties and deadlines. Output as Markdown."
- The structured result is embedded in a minutes template.
- Optional: automated delivery to all participants; ticket creation in your project management tool.
The full pipeline runs on your own infrastructure, no cloud, no API costs, no vendor lock-in.
New competition: Voxtral and other entrants
Whisper is no longer the only serious player. In February 2026, Mistral AI released Voxtral Transcribe 2, a second-generation speech-to-text family comprising Voxtral Mini Transcribe V2 for batch transcription and Voxtral Realtime for live streaming. Based on Mistral's published figures and early community comparisons, Voxtral targets strong multilingual performance.
The critical difference for SMBs: Voxtral is primarily a cloud API product. Whisper, and its faster-whisper and whisper.cpp variants, remains the only mature option that runs fully on-premise without any cloud connection.
Other open-source entrants in 2025-2026, including Canary and Granite speech models, are gaining ground in English benchmarks but have narrower multilingual coverage than Whisper's 99-language support. For European multilingual teams, Whisper large-v3 and turbo remain the practical choice.
Why local transcription is the right GDPR move
Meeting audio regularly contains personal data under GDPR: employee names, salary figures, health disclosures, client details. Sending that to any cloud service requires:
- A lawful basis (Art. 6 GDPR) for the transfer
- A signed data processing agreement (Art. 28 GDPR) with the vendor
- For US vendors: additional transfer safeguards (Art. 46 GDPR)
Local processing eliminates all of this. Data never leaves your network. There is no adequacy decision to rely on, no DPA to negotiate, no risk from future legal challenges like a hypothetical Schrems III.
EU AI Act Art. 50 adds a transparency requirement: AI-generated summaries that "represent the views or statements of real persons" must be labelled as AI-generated. That is simple to handle internally, and far easier to document when the entire pipeline is under your own control. Our full treatment of these obligations is on our data sovereignty page.
Practical applications
Law firm (12 staff): Client consultations transcribed and filed as case notes without any client data leaving the firm's server.
Manufacturing plant (90 employees): Weekly production meetings auto-transcribed; action items pushed directly to Jira without manual follow-up.
HR team (mid-sized company): Job interviews transcribed locally; no cloud service accesses applicant data; data minimisation under Art. 5 GDPR is technically enforced, not just documented.
Consultancy (multi-language): Client meetings in English, German, and Spanish auto-transcribed. Whisper's multilingual model switches between languages per recording without additional configuration.
Cost profile
Running Whisper locally has no per-minute API cost. The hardware investment, a Mac Studio M3 Ultra at approximately €4,000-5,000, or a refurbished Linux server with an NVIDIA GPU at a similar price point, pays for itself against cloud transcription costs for a team that records 20+ hours of meetings per month, based on typical reported cloud pricing.
For teams already running a local AI stack for LLM inference, adding Whisper to the same hardware adds no significant marginal cost.
Getting started
If you want to run meeting transcription entirely on-premise, or connect a transcription pipeline to your existing workflows through our infrastructure, start a pilot project with Freshlab: 14 days, fixed scope, a clear go/no-go at the end.
Questions? Contact us directly. We work in English, German, and Spanish.