faster-whisper
Whisper, 4× faster, same accuracy. CTranslate2 backend.
Open Source 2–6 GB VRAM
- Min VRAM
- 2 GB
- GPU class
- Entry GPU
- Quant
- INT8
Actually FreeNo SignupOpen SourceWatermark-Free
OpenAI Whisper
The reference open-source speech-to-text model.
Open Source 2–10 GB VRAMRuns locally
Actually FreeNo SignupOpen SourceWatermark-FreeHobbyist-FriendlyAPI
Visit OpenAI WhisperUpdated 2025-11-15 · Direct link
Hardware requirements
Runs locally · Entry GPU (6–8 GB)
Min VRAM
2 GB
Rec. VRAM
10 GB
Min RAM
8 GB
Rec. RAM
16 GB
Disk
10 GB
GPU class
Entry GPU
11.3+Apple Silicon ✓CPU-CapableQuant: FP16
`tiny` runs on CPU; `large-v3` needs ~10 GB VRAM.
Screenshot placeholder · OpenAI Whisper
What is OpenAI Whisper?
Whisper is the open-weight transcription model that reset expectations for ASR. Five sizes (tiny → large-v3), 99 languages, robust to noise and accents, runs fully offline. The reference implementation is slow; in practice you'll want faster-whisper, whisper.cpp, or WhisperX. Listed here as the canonical entry.
Pros & cons
Pros
- Multilingual ASR that genuinely works across 99 languages
- Five size points let you trade speed for accuracy
- Active fork ecosystem (faster-whisper, WhisperX, whisper.cpp)
Cons
- Reference impl in PyTorch is 4-6× slower than the optimised forks
- Hallucinates plausible text during long silences
What's actually free?
MIT, weights freely downloadable from HuggingFace.
✓ 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