Meeting-note SaaS sends confidential business discussions to third-party US infrastructure. A self-hosted pipeline on the RTX 5060 Ti 16GB at our UK dedicated GPU hosting processes an hour of audio in roughly 100 seconds end to end (transcribe, diarise, summarise, extract actions) and keeps the recording inside your network. The Blackwell card’s 16 GB GDDR7 and native FP8 are enough to hold Whisper, pyannote and Llama 3 8B concurrently.
Contents
- Pipeline stages and timings
- Zoom and Teams ingestion
- Summary format
- Cost vs SaaS meeting notes
- Privacy and retention
Pipeline
| Stage | Tool | Time per hour of audio |
|---|---|---|
| Download recording | Zoom / Teams webhook | ~5 s (depends on link) |
| Transcription | Whisper Turbo (faster-whisper, FP16) | ~60 s |
| Diarisation | pyannote.audio 3.1 | ~25 s |
| Summary + actions | Llama 3.1 8B FP8 | ~10 s (2k input, 500 out) |
| Embedding for search | BGE-base | <1 s |
| Total | ~100 s |
Integrations
- Zoom: Webhook
recording.completed-> download MP4 via signed URL -> feed pipeline - Microsoft Teams: Graph API subscription on
CallRecordor SharePoint recording folder, pull via Graph - Google Meet: Drive API watch on the meeting-recordings folder
- Manual upload: Web UI for MP3, MP4, M4A, WAV up to 4 hours
- Live capture: LiveKit/Meet agent records and streams audio; pipeline runs continuously
Output
- Clean transcript with speaker labels and timestamps
- Executive summary (5-10 sentences)
- Bulleted action items with owners and due dates
- Decision log with rationale
- Open questions and follow-ups
- Sentiment markers (optional, off by default)
Cost
| Team profile / month | Otter.ai Business | Fireflies Pro | Self-hosted 5060 Ti |
|---|---|---|---|
| 10 users, 40 h/user | ~£200 | ~£180 | Flat £300 (unlimited) |
| 50 users, 30 h/user | ~£1,000 | ~£900 | Flat £300 |
| 200 users, 20 h/user | ~£4,000 | ~£3,600 | Flat £300-450 (one box) |
Beyond roughly 20 active users the dedicated box is cheaper and faster. One 5060 Ti handles 36 hours of meeting audio per hour of wall time, so even 1000 hours/month finishes in under 30 hours of GPU wall time.
Privacy
Recordings, transcripts and summaries live on your UK infrastructure. No third-party sub-processor agreements are required in customer DPAs. Retention is whatever your policy says. Legal, healthcare and finance teams that were previously blocked from using SaaS meeting notes by their compliance team can usually adopt a self-hosted equivalent without friction.
Private meeting notes in 100 seconds per hour
Whisper plus Llama on Blackwell 16GB. UK dedicated hosting.
Order the RTX 5060 Ti 16GBSee also: Whisper benchmark, webinar transcription, internal tooling, document Q&A, Llama 3 8B benchmark.
