Apple's SpeechAnalyzer Beats Whisper Small in On-Device Speech Test
Benchmark shows Apple's SpeechAnalyzer outperforms Whisper Small and the legacy SFSpeechRecognizer on LibriSpeech word error rate and speed.
An independent benchmark measured Apple's new SpeechAnalyzer API, introduced in iOS/macOS 26, against three Whisper models and the legacy SFSpeechRecognizer using LibriSpeech's clean and noisy test splits. Every engine ran fully on-device on the same Apple Silicon machine through identical production code paths.
The result is unambiguous: SpeechAnalyzer produced a lower word error rate than even the largest Whisper model tested, Whisper Small, on both splits, while running roughly three times faster. The legacy SFSpeechRecognizer, by contrast, trailed even the tiny 40MB Whisper model on clean speech, making the case for migration straightforward.
The methodology is notably transparent: the Whisper measurements closely reproduce OpenAI's own published numbers, raw per-utterance transcripts were released publicly, and the testing process itself surfaced a bug in the tester's own app caused by a missing finalization call in the SpeechAnalyzer integration.
The findings are limited to English and read-speech audio on a single chip, so Whisper remains relevant for multilingual or meeting-style audio. But for English on-device transcription, Apple's built-in engine is now the strongest option measured.