Apple SpeechAnalyzer Benchmarked Against Whisper
Original: Apple's new SpeechAnalyzer API, benchmarked against Whisper and its predecessor
Why This Matters
Apple's built-in speech engine now outperforms Whisper Small on accuracy and speed, reshaping on-device ASR choices for Apple developers.
Inscribe tested Apple's new SpeechAnalyzer API (iOS/macOS 26) against SFSpeechRecognizer and three Whisper models on 5,559 LibriSpeech utterances using an Apple M2 Pro. SpeechAnalyzer achieved 2.12% WER on clean speech, outperforming Whisper Small (3.74%) while running ~3x faster.
AI workspace app Inscribe conducted the first published benchmark of Apple's SpeechAnalyzer, the on-device speech recognition API introduced in iOS 26 and macOS 26 to replace SFSpeechRecognizer. Apple shipped the new API with no official accuracy figures, leaving developers without data to guide migration decisions.
Testing on an Apple M2 Pro (32GB, macOS 26.5.1) with the LibriSpeech dataset (5,559 total utterances across clean and noisy splits), results showed SpeechAnalyzer achieved a 2.12% word error rate (WER) on test-clean and 4.56% on test-other. Whisper Small followed at 3.74% and 7.95%, while the legacy SFSpeechRecognizer scored 9.02% and 16.25% respectively — placing last on clean audio, behind even Whisper Tiny (40MB).
SpeechAnalyzer also ran approximately three times faster than Whisper Small per second of audio. All engines exceeded real-time processing speed, transcribing audio at 12x–40x real time on the M2 Pro.
Whisper retains advantages in language coverage (~30 locales vs. SpeechTranscriber's broader Whisper range) and cross-platform availability. However, for English on Apple hardware, Inscribe concluded SpeechAnalyzer is now the strongest on-device option and updated its own product defaults to prefer SpeechAnalyzer for supported languages.