Transcribe on the GPU when there is one, fall back to the CPU

[simplaerai-sv]

What

scripts/transcribe.py runs faster-whisper on the GPU when one is
visible, and keeps today's CPU path as the fallback. One file changed.

Why

The runner hard-codes device="cpu", compute_type="int8". On a machine
with an NVIDIA card that leaves most of the box idle. Measured on the same
60-second Ukrainian clip, RTX 4060 (8 GB), 12-core CPU:

model	CPU int8	GPU float16
small	15.5 s	6.9 s
large-v3	—	8.8 s (incl. model load)

A 50-minute lecture on large-v3 goes from "leave it running" to about
seven minutes.

How

• pick_device() — cuda when ctranslate2 reports a CUDA device, else
  cpu. KADR_WHISPER_DEVICE=cpu|cuda|auto overrides (default auto),
  so a user can pin the CPU without touching code.
• load_model() — cuda/float16, falling back to cpu/int8 if the load
  itself fails (driver too old, no VRAM).
• The transcribe loop moved into run(). If the GPU fails during
  inference — the common Windows case is a missing cublas64_12.dll,
  the model loads fine and the first kernel dies — the job is redone on
  the CPU. Only while nothing has been streamed yet: a CPU rerun after
  partial GPU output would hand the editor every segment twice.
• _register_nvidia_dlls() (Windows only) — the pip wheels
  nvidia-cublas-cu12 / nvidia-cudnn-cu12 drop their DLLs under
  site-packages/nvidia/<lib>/bin, which is not on the search path;
  registering those folders is what makes CUDA actually work from a plain
  pip install. No-op on Linux/macOS.
• The chosen device is logged to stderr; stdout stays NDJSON-only, so
  electron/transcribe.ts needs no change.

Behaviour on a machine without a GPU is byte-for-byte what it was.

Tested

Windows 11, Python 3.12, faster-whisper 1.2.1, ctranslate2 4.8.2:

• GPU path: small and large-v3, timings above, Ukrainian recognised
• inference-level fallback: with the CUDA wheels absent, cuda load →
  cublas64_12.dll error → CPU rerun → exit 0, all segments present
• KADR_WHISPER_DEVICE=cpu → CPU path, unchanged output
• device-less path not testable here; it is the untouched original code

Users on Windows who want the GPU need
pip install nvidia-cublas-cu12 nvidia-cudnn-cu12 on top of
faster-whisper — happy to add a line to the README if you would like it
in this PR.

🤖 Generated with Claude Code