What is Whisper?
Whisper is OpenAI's Speech-to-Text model – completely open source and runnable locally. It transcribes audio in over 90 languages with impressive accuracy.
Unlike cloud services, you don't pay API costs and your audio data stays private. Perfect for confidential interviews, customer calls, or sensitive meeting recordings.
Whisper vs. Cloud Services
Google Speech-to-Text costs $0.006 per 15 seconds. For 10 hours of podcast per month, that's ~$15. A VPS for €5/month transcribes unlimited – and your data never leaves your server.
For GDPR-sensitive content (interviews, customer conversations, patient data), self-hosting is the only safe option. No third parties, no data sharing.
| Service | Cost/Month (10h Audio) | Privacy |
|---|---|---|
| Google Speech-to-Text | ~$15 | US Cloud |
| AWS Transcribe | ~$14 | US Cloud |
| Whisper API (OpenAI) | ~$6 | US Cloud |
| Whisper Self-Hosted | ~€5 VPS | 100% local |
faster-whisper: The Better Alternative
faster-whisper uses CTranslate2 and is 4x faster than the original with the same accuracy. On a 4-core VPS, you can transcribe 1 hour of audio in about 15-20 minutes instead of 60+ minutes.
Installation via Docker: A single docker-compose.yml is all you need for a production-ready transcription setup with REST API.
Hardware Requirements by Model
RAM requirements depend on the chosen model. For most languages, the 'Small' model offers the best balance of speed and accuracy.
| Model | RAM | Duration (1h Audio) | Quality |
|---|---|---|---|
| Tiny | 2 GB | ~8 min | Sufficient |
| Base | 2 GB | ~15 min | Good |
| Small | 4 GB | ~20 min | Very good |
| Medium | 8 GB | ~40 min | Excellent |
| Large-v3 | 12 GB | ~60+ min | Best |
Our Recommendation
For occasional transcription with the Small model, Hetzner CX22 with 4 GB RAM for about €4/month is sufficient.
For regular use or the Medium model, we recommend Contabo Cloud VPS S with 8 GB RAM for about €6/month.
For batch processing large audio volumes: Contabo Cloud VPS M with 16 GB RAM for about €10/month.
