Whisper audio encoders compiled for the Apple Neural Engine (ANE) with
ANEForge, for use with the
whisper-aneforge backend
for
whisper.cpp.
Each bundle is self-contained in its own directory. The ANE program is
compiled on-device at load (about 0.6 s) from model.mil + weights.bin;
no prebuilt binaries are shipped.
Download a bundle (e.g. whisper-small-ane):
Build whisper.cpp with the ANEForge backend (see the whisper-aneforge
README), then run, pointing the backend at the bundle and the ANEForge
dispatch dylib:
1ANEFORGE_DYLIB=$(python -c "import aneforge,os;print(os.path.join(os.path.dirname(aneforge.__file__),'_lib','libane_e5rt_dispatch.dylib'))") \
2ANEFORGE_ENCODER=models/whisper-small-ane \
3 ./build/bin/whisper-cli -m models/ggml-small.bin -f samples/jfk.wav
