whisper.cpp
models/generate-coreml-model.sh large-v3 --encoder-only True
time ./main -m models/ggml-large-v3.bin -f test.wav -l auto
| Neon & MPS 👍 | CoreML 🚀 (47%) | |
|---|---|---|
| load time | 1007.19 ms | 859.73 ms |
| mel time | 216.87 ms | 224.71 ms |
| sample time | 3550.35 ms | 2659.66 ms |
| encode time | 7821.69 ms | 5801.61 ms |
| decode time | 2958.22 ms | 4105.18 ms |
| batchd time | 88241.95 ms | 54016.19 ms |
| prompt time | 1618.32 ms | 1218.58 ms |
| total time | 105432.62 ms | 71318.02 ms |
| cpu time | 1:45.50 | 1:11.44 |
速度提高了,但效果下降了。