Apple SpeechAnalyzer beats small Whisper models on device in a LibriSpeech benchmark

Category: Apple
Status: discussion
Sources: benchmark, HN 48894752, heise
Summary: A benchmark published 2026-07-13 ran Apple's SpeechAnalyzer API on an M2 Pro under macOS 26.5.1 against Whisper and the legacy SFSpeechRecognizer over 5,559 LibriSpeech utterances. SpeechAnalyzer reported 2.12% word error rate on clean speech and 4.56% on noisy speech, ahead of Whisper Small (3.74% and 7.95%) and the legacy API (9.02% and 16.25%), while running about 3x faster than Whisper Small per second of audio. The author notes larger Whisper models such as Large V3 Turbo still lead on accuracy, so the on-device engine is a speed-for-accuracy tradeoff.
Why it matters: On-device transcription on current macOS now matches or beats small Whisper models with no model download and no server round trip, changing the build-versus-bundle choice for local speech features.