Just convert documents
1 vCPU · 2 GB RAM · ~5 GB disk. A cheap VPS is plenty. No anonymization service needed.
System requirements
Let’s be honest up front, so you never deploy Escriba and feel it “doesn’t work” because of your hardware. The base app is light and runs on a modest VPS. The heavy parts — enterprise PII anonymization, and large-model audio transcription — are optional and only ask for more when you actually turn them on.
The base app — light by design
Section titled “The base app — light by design”Document conversion (PDF, Word, Excel, images, OCR for normal scans) is fast and frugal. Escriba spawns one worker per CPU core, and each worker uses ~250 MB of RAM, so it adapts to whatever host you give it.
| Minimum | Recommended | |
|---|---|---|
| CPU | 1 vCPU | 2 vCPU |
| RAM | 2 GB | 4 GB |
| Disk | ~5 GB (image is ~3 GB) | ~6 GB on SSD |
| Good for | Personal use, light documents | A small team, smooth OCR |
Enterprise PII anonymization — an optional module
Section titled “Enterprise PII anonymization — an optional module”Escriba’s anonymization doesn’t run inside the main app. It lives in a separate, internal-only service — Anonimal — that you mount only if you need it. That’s a feature, not a compromise: the privacy engine is isolated, it never has to be exposed, and your light converter stays light when you don’t need it.
It’s built for serious, institutional security — the kind of setting where confidential documents simply cannot leave your infrastructure. Because it loads a full NER model, it is heavier:
| With anonymization (Escriba + Anonimal) | |
|---|---|
| CPU | 2 vCPU min · 4 vCPU recommended |
| RAM | 6 GB min · 8 GB recommended (the model holds ~3 GB resident) |
| Disk | ~12 GB (Anonimal’s image is ~7 GB on top of Escriba’s) |
| Good for | Companies, public bodies, anyone who can’t use a third-party cloud |
Audio, video & heavy OCR — CPU-bound
Section titled “Audio, video & heavy OCR — CPU-bound”Transcription (Whisper) and OCR run on the CPU in the bundled image — there’s no
GPU requirement, but bigger Whisper models are slower and hungrier. Pick the model
that fits your hardware with WHISPER_MODEL:
| Whisper model | Extra RAM (approx.) | Notes |
|---|---|---|
tiny / base (default) | ~1 GB | Fast; fine for clear speech on modest hardware |
small | ~2 GB | A good accuracy/speed balance |
medium | ~5 GB | Noticeably slower on CPU |
large-v3 | ~10 GB | Most accurate; needs a strong server, slow on CPU |
Which setup is right for you?
Section titled “Which setup is right for you?”Add enterprise PII privacy
2–4 vCPU · 6–8 GB RAM · ~12 GB disk. Mount the Anonimal module for institutional-grade redaction.
Heavy transcription
4+ vCPU · 8–16 GB RAM. For large Whisper models and lots of audio/video.