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
WhisperX
Whisper + speaker diarisation + word-level timestamps.
Open Source 4–8 GB VRAMRuns locally
Actually FreeOpen SourceWatermark-FreeHobbyist-Friendly
Visit WhisperXUpdated 2026-03-12 · Direct link
Hardware requirements
Runs locally · Entry GPU (6–8 GB)
Min VRAM
4 GB
Rec. VRAM
8 GB
Min RAM
16 GB
Rec. RAM
16 GB
Disk
12 GB
GPU class
Entry GPU
11.8+No Apple SiliconGPU RequiredQuant: INT8, FP16
Diarisation model adds ~2 GB VRAM on top of Whisper.
Screenshot placeholder · WhisperX
What is WhisperX?
WhisperX takes faster-whisper and adds the things production transcription actually needs: forced-alignment for word-accurate timestamps (via wav2vec2), speaker diarisation (via pyannote), and VAD-based chunking. The de facto open-source pipeline for subtitling and podcast transcripts.
Pros & cons
Pros
- Word-level timestamps that are actually word-level
- Speaker diarisation in the same pipeline
- Built on faster-whisper — speed inherited
Cons
- Diarisation step needs accepting HF model terms
- Heavier dependency stack (PyAnnote, wav2vec2)
What's actually free?
BSD-4. Note: pyannote diarisation requires HF model access (free).
✓ Actually FreeOpen SourceWatermark-Free
Alternatives
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