Originally reported as NixOS/nixpkgs#409684, but then, during bisection, discovered that it's caused by an upstream change.

Bisection led me to commit 21d890d.

Reproduction

1. cmake -B build -GNinja -DGGML_BACKEND_DL=TRUE
2. ninja -C build
3. build/bin/whisper-cli -m ~/ggml-base.en.bin -f samples/jfk.wav
4. Observe a segfault caused by the most spectacular null pointer dereference I've ever seen 🤯
5. rm -rf build (just in case)
6. git revert --no-edit 21d890d53441a08e36992b0f68221f38b715789d
7. cmake -B build . -GNinja -DGGML_BACKEND_DL=TRUE && ninja -C build && build/bin/whisper-cli -m ~/ggml-base.en.bin -f ./samples/jfk.wav
8. Observe Whisper magically working again ✨

Notes

Well, obviously, reverting a commit that, on the other hand, improves Whisper.cpp's device support is a no-no. (This commit is part of #2863, which claims to expand aarch64 support.) But this clearly seems to be a bug.

Workarounds

Uh, don't build with the DL backend. But then I assume whisper.cpp gets locked into the architecture of the machine it was built on, whereas with the DL backend, one could use another CMake option to build shared libraries for several CPU types at once and, supposedly, choose the right one at runtime.