Apple’s new Speech Analyzer API is the most accurate on‑device speech engine we’ve tested, surpassing every Whisper model we ship. On the LibriSpeech test‑clean set it achieved a 2.12% word error rate (WER) versus 3.74% for Whisper Small, and on the noisy test‑other set it posted 4.56% WER compared to Whisper Small’s 7.95%. It also runs about three times faster than Whisper Small, processing an hour of audio in roughly 1.5–5 minutes on an M2 Pro.

The legacy SFSpeech Recognizer fell far behind, with 9.02% clean‑speech WER and 16.25% noisy‑speech WER. Migrating to Speech Analyzer cuts errors by 3.5–4× and adds punctuated, cased output. While Whisper still supports ~30 languages and runs on non‑Apple platforms, for English transcription on current iPhone or Mac hardware the built‑in engine is now the clear winner.

Our benchmark ran all five engines through identical production code on the same device, using the same LibriSpeech corpus and normalization pipeline. Results validate Whisper’s published numbers and provide transparent raw transcripts for independent verification.