Less hacky integration with whisper.cpp

[shun-liang]

On Apple Silicon, whisper.cpp runs much faster than faster-whisper, as the whisper.cpp accelerates with the Apple GPU through CoreML, while faster-whisper only supports running on the CPUs on Apple Silicon.

Faster-whisper relies on CTranslate2 for Transformers inferences. There does not seem to be any hope that CoreML will be supported by that project. (See OpenNMT/CTranslate2#1607 and OpenNMT/CTranslate2#1586)

Right now, yt2doc uses faster-whisper as the transcription backend by default, but does also support whisper.cpp. The whisper.cpp support however is somewhat hacky (see here) and cumbersome as it requires the user to have installed/compiled whisper.cpp on their device themselves.

It's possible to use whisper.cpp through one of its Python bindings. However, among them, only pywhispercpp is actively maintained and claims to have supported CoreML. The CoreML support of pywhispercpp, however, requires cloning the repository and build the project locally with an environment variable that feeds into the building process. I have not found a way to include pywhispercpp with the environment variable in the pyproject.toml of this project as a dependency, which is essntial to distribute yt2doc through PyPI.

Any idea or solution on this is much appreciated.

[shun-liang]

Alternatiely, we may be able to drop both faster-whisper and whisper.cpp if running Whisper models on Hugging Face's transformers gives good support on Apple GPU without faff. Need to investigate that too.