Apple's SpeechAnalyzer beats Whisper Small with 44% lower error rate

Apple has quietly replaced its SFSpeechRecognizer with a new API, SpeechAnalyzer (and SpeechTranscriber), in iOS 26 and macOS 26. A thorough benchmark from Inscribe, a private on-device AI workspace, tested all available Apple and Whisper models on the standard LibriSpeech dataset. The results are stark: SpeechAnalyzer achieved a 2.12% word error rate (WER) on clean speech and 4.56% on noisy speech, compared to 3.74% and 7.95% for Whisper Small, the largest Whisper model tested. The legacy SFSpeechRecognizer trailed badly at 9.02% and 16.25%, respectively.

SpeechAnalyzer also proved roughly 3x faster than Whisper Small per second of audio, all running on-device on an M2 Pro Mac. Inscribe's director notes that for English transcription on current iPhones and Macs, the built-in engine is now the strongest on-device option. The Whisper models retain advantages in language coverage (30+ locales vs. about 30 for SpeechTranscriber) and platform independence. Developers currently using SFSpeechRecognizer should migrate immediately—the new API offers a 3.5–4x accuracy improvement on the same audio. Inscribe has already updated its own Auto engine to prefer SpeechAnalyzer for supported languages.