faster-whisper on Apple Silicon CPU: beam_size=1 is ~45% faster; cpu_threads=all cores is 2-3x SLOWER
Measured with faster-whisper 'small', device='cpu', compute_type='int8', word_timestamps=True on a 10-core M-series Mac (Nov 2026). 83 s English clip: default (beam 5, cpu_threads=0) 15.7-17.4 s; beam_size=1 9.0-10.8 s; cpu_threads=10 with beam 5 50.0 s (!); cpu_threads=10 with beam 1 21.9 s; 'base' model beam 5 6.6 s. Setting cpu_threads to os.cpu_count() pulls in the efficiency cores and makes CTranslate2 much slower, so leave the default. Greedy vs beam-5 word overlap (difflib on word lists incl. punctuation): English 97% (differences mostly punctuation), Spanish 203 s clip 78% (punctuation plus some word changes), so check non-English output before switching to greedy. Model load is only ~0.7 s and audio extraction via ffmpeg to 16 kHz f32 ~0.1 s; the decode itself is ~90% of the stage. faster-whisper has no Metal/MPS backend; for GPU on a Mac you need a different implementation (e.g. mlx-whisper or whisper.cpp).