Apple's SpeechAnalyzer Beats Whisper on Accuracy in First Independent Benchmark

Inscribe benchmarked Apple's new on-device SpeechAnalyzer API against Whisper on 5,559 LibriSpeech utterances — it scored 2.12% WER, beating Whisper Small while running 3x faster.

Apple quietly shipped a new speech recognition engine with iOS 26 and macOS 26 — and it turns out to be the best on-device option for English. The first independent benchmark, published by transcription app Inscribe, puts Apple's SpeechAnalyzer ahead of every Whisper model they tested.

The numbers: SpeechAnalyzer scored a word error rate (WER) of 2.12% on clean speech and 4.56% on noisy speech, compared to 3.74% and 7.95% for Whisper Small. The old SFSpeechRecognizer API it replaces came in at 9.02% — nearly four times worse. SpeechAnalyzer also runs about three times faster than Whisper Small on an M2 Pro.

All five engines ran fully on-device. No audio leaves the device, which means the private transcription option is no longer the accuracy compromise it used to be. Inscribe changed its own product defaults based on these results: SpeechAnalyzer now handles supported languages, with Whisper as the fallback for everything else.

The catch: SpeechAnalyzer supports around 30 languages, while Whisper covers far more. For multilingual users or non-Apple platforms, Whisper remains essential. But for English on a current iPhone or Mac, the best engine is already built in.

Commenters on Hacker News noted the benchmark's methodology is unusually rigorous — the authors released raw transcripts so anyone can independently rescore the results.

Sources: Inscribe, Hacker News