OpenAI Whisper
The reference open-source speech-to-text model.
Open Source 2–10 GB VRAM
- Min VRAM
- 2 GB
- GPU class
- Entry GPU
- Quant
- FP16
Actually FreeNo SignupOpen SourceWatermark-Free
faster-whisper
Whisper, 4× faster, same accuracy. CTranslate2 backend.
Open Source 2–6 GB VRAMRuns locally
Actually FreeNo SignupOpen SourceWatermark-FreeHobbyist-Friendly
Visit faster-whisperUpdated 2026-04-25 · Direct link
Hardware requirements
Runs locally · Entry GPU (6–8 GB)
Min VRAM
2 GB
Rec. VRAM
6 GB
Min RAM
8 GB
Rec. RAM
16 GB
Disk
8 GB
GPU class
Entry GPU
11.8+Apple Silicon ✓CPU-CapableQuant: INT8, FP16
`large-v3` runs in 4-5 GB VRAM with INT8.
Screenshot placeholder · faster-whisper
What is faster-whisper?
faster-whisper reimplements Whisper inference on top of CTranslate2 (C++/CUDA), delivering 4× speedup over the reference PyTorch impl at the same word error rate. INT8 quantisation halves VRAM again with no measurable accuracy loss. The default Whisper backend for anyone who's measured it.
Pros & cons
Pros
- 4× faster than reference Whisper at equal accuracy
- INT8 quantisation cuts VRAM in half
- Drop-in CLI compatible with the reference
Cons
- No diarisation built in — pair with WhisperX or pyannote
- Setup involves CUDA + cuDNN library paths that occasionally fight
What's actually free?
MIT.
✓ Actually FreeNo SignupOpen SourceWatermark-Free
Alternatives
WhisperX
Whisper + speaker diarisation + word-level timestamps.
Open Source 4–8 GB VRAM
- Min VRAM
- 4 GB
- GPU class
- Entry GPU
- Quant
- INT8
Actually FreeOpen SourceWatermark-FreeHobbyist-Friendly