RTX 3060 server: private call transcription with faster-whisper
Audio uploaded from a phone PWA is transcribed on a local RTX 3060 (12GB) with faster-whisper int8, sharing the card with an embedding model and Gemma.
Solo2-10 peopleTranscriptionUnder $1kSensitivity: High (client/confidential)
Hardware
| Item | Approx. price (as of 2026-10-16) |
|---|---|
| Server with NVIDIA GeForce RTX 3060 (12GB) | Not stated in source |
| NVIDIA GeForce RTX 3060 12GB (various) (Newegg, related current price) | $449.99 |
Source prices are as stated by the source. Retailer offers: prices checked 2026-10-16, may change. Some links may be affiliate links; we may earn a commission at no extra cost to you. This never changes what we list.
Software stack
- faster-whisper (int8)
- Custom PWA for uploads
Models used
- Whisper medium (int8, ~2.5-3GB VRAM)
- bge-m3
- Gemma
Good at
- Audio never leaves the machine
- 8-12s per minute of audio on GPU
Weak at
- Falls back to CPU under VRAM pressure: 30-60s per minute
- Names and cross-talk degrade accuracy
Rough total cost
Not stated in source.
Getting started
- Install NVIDIA drivers + CUDA on the server.
- pip install faster-whisper and load the medium model with compute_type int8.
- Provide an upload path (the author uses a PWA).
- Load models sequentially (transcribe, embed, analyze) to fit 12GB.
Source & credit
Why I run speech-to-text locally instead of calling a cloud API — DEV Community (hannune)
Summarized from the source; quotes are the source author's words. Check the original before buying.