Allow building of whispercpp with COREML support

[tangm]

Addresses #12 and maybe #20

Rather than copying much of whisper.cpp's CMakeLists.txt, just link the built libraries against main.cpp.

To enable CoreML support on macs per https://github.com/ggerganov/whisper.cpp?tab=readme-ov-file#core-ml-support , we will still need to first download and convert the appropriate model using the original whisper.cpp repository, producing a <model>.mlmodelc directory.

To build and install,

• export CMAKE_ARGS="-DWHISPER_COREML=1"
• python -m build --wheel in this repository to build the wheel. Assumes you have installed build with pip install build
• pip install dist/<generated>.whl

To invoke,

• Use something like model = Model('<model_path>/ggml-base.en.bin', n_threads=6), assuming the converted mlmodelc directory is also in <model_path>

If successful, you will see something like:

whisper_init_state: loading Core ML model from '<model_path>/ggml-base.en-encoder.mlmodelc'
whisper_init_state: first run on a device may take a while ...
whisper_init_state: Core ML model loaded

You can also verify if whisper.cpp has been built successfully by doing print(Model.system_info()), and you should see COREML = 1

[absadiki]

Thanks @tangm very much for the PR.
Yes I agree, your idea is perfect.

I don't have access to a MAC to test, but the mac-os workflow has been successfully finished.

Could you please try to add "-DWHISPER_COREML=1" to the cmake args in the setup.py file and test if you can build the project with pip install -e . ? so we can generate pre-built wheels for pypi.

[tangm]

Thanks @tangm very much for the PR. Yes I agree, your idea is perfect.

I don't have access to a MAC to test, but the mac-os workflow has been successfully finished.

Could you please try to add "-DWHISPER_COREML=1" to the cmake args in the setup.py file and test if you can build the project with pip install -e . ? so we can generate pre-built wheels for pypi.

hmm, I can add it to the build args in setup.py, but I'm not sure if you want to turn on CoreML support by default, as it does require some intervention by the user to convert models, i.e. someone on a mac who has been using this library previously but upgrades will now experience a crash because whisper.cpp with CoreML support expects converted models to be present. Maybe just putting the instructions in the README like whisper.cpp did?

If there is some way of releasing variations of a library that might be more suitable but I'm not very familiar with the pypi ecosystem...

[absadiki]

Yeah .. And also I forget that the Github VMs do not have CoreML to build the binary against (or at least AFAIK), so instructions on the README file should be enough.

Thank you very much again, @tangm, for the awesome contribution.

[tangm]

Glad to help! 👍