mlx-whisper vs faster-whisper on an M-series Mac: same 'small' model 2-4x faster on GPU; large-v3-turbo ~same time as CPU small
Measured Nov 2026, word_timestamps=True, warm (second) runs, 16 kHz float32 numpy audio from ffmpeg. 83 s English clip: faster-whisper small int8 CPU 19.2 s; mlx-community/whisper-small-mlx 4.4 s; mlx-community/whisper-large-v3-turbo 14.5 s. 203 s Spanish clip: 36.9 s / 17.3 s / 41.5 s. In a real pipeline run, mlx small took 3.8 s for 76 s of English and 5.6 s for 87 s of Spanish. Output: mlx_whisper.transcribe(audio_np, path_or_hf_repo=repo, word_timestamps=True) returns segments[i]['words'] with word/start/end (use segment.get('words', []) for silent segments) and result['language']. Quality vs faster-whisper small beam 5: mlx small differs mostly in punctuation/capitalisation (94% word match English, 81% Spanish), similar to greedy decoding; turbo got proper names right (e.g. 'Noem' where small wrote 'Noam') and used Spanish inverted question marks. Install with uv pip install mlx-whisper; the first call downloads the model from Hugging Face (set HF_HUB_DISABLE_XET=1 if the xet download errors). Apple Silicon only. Worked fine in a subprocess running at the same time as a separate torch MPS process (Florence-2 captions).