gpu-setup.md

GPU setup (optional — for fast large-v3 transcription)

faster-whisper runs large-v3 on an NVIDIA GPU via CTranslate2. On the target machine (RTX 3060 Laptop 6GB, Windows 11) transcription runs at roughly RTF 0.45 (~80 min for a 3-hour meeting; faster on real meetings once silence is skipped).

Without a GPU, the pipeline still works — it falls back to CPU automatically (device="cpu", compute_type="int8"), just slower.

Enable large-v3

The pipeline defaults to the small model (WHISPER_MODEL in .env, see src/config.py) so it stays fast on CPU-only machines out of the box. To actually get the large-v3 quality/GPU benefit described above, you must opt in by setting WHISPER_MODEL=large-v3 in your .env. On GPU this is fast (~RTF 0.45, as above). On CPU-only machines large-v3 is very slow (~RTF 3.8) — if you don't have a GPU, leave WHISPER_MODEL set to small.

One-time install (Windows, CUDA GPU)

CUDA-enabled PyTorch (used for GPU detection and pyannote), matching the driver's CUDA version (13.0 on the target machine):

.\.venv\Scripts\python -m pip install --index-url https://download.pytorch.org/whl/cu130 "torch==2.13.0+cu130" "torchaudio==2.11.0+cu130"

CUDA 12 cuBLAS + cuDNN 9 that CTranslate2 loads at runtime (these ship the cublas64_12.dll / cudnn_ops64_9.dll that ctranslate2 needs; PyTorch's own CUDA 13 libraries are not compatible with ctranslate2's CUDA 12 build):

.\.venv\Scripts\python -m pip install nvidia-cublas-cu12 nvidia-cudnn-cu12

src/transcribe.py registers these libraries' bin directories as DLL search paths automatically at model-load time, so no PATH changes are needed.

Verify

.\.venv\Scripts\python -c "import torch; print('cuda:', torch.cuda.is_available())"

Should print cuda: True. If it prints False, transcription falls back to CPU.