Hi everyone and thank you for reading my PR !

Here's a summary of what it does :

• Override extractor devices to CPU via config when CUDA is unavailable, so extractors are built with the correct device from the start
• Fix whisperx (ctranslate2) compatibility: fall back to CPU with int8 compute type since it does not support MPS
• Add "Running on Apple Silicon" section to README with working example including the required multiprocessing.set_start_method("fork") workaround

Here's the test plan :

• Tested end-to-end inference on Apple Silicon M4 (macOS) with a video file
• Verified feature extraction (LLaMA 3.2, V-JEPA2, Wav2Vec-BERT) and brain prediction complete successfully
• Verified plotting output (brain_activity.png) renders correctly
• Verify no regression on Linux/CUDA: all changes check torch.cuda.is_available() first, so CUDA environments are unaffected

Tested on MacBook Pro M4 Pro