Apple's SpeechAnalyzer beats Whisper Small on-device

Apple's SpeechAnalyzer API out-transcribes Whisper Small on clean English audio, per a benchmark from get-inscribe. The word error rate came in at 2.12%, against 3.74% for Whisper Small and 9.02% for Apple's older SFSpeechRecognizer, a 43% cut in errors. It also ran about three times faster than Whisper Small on an M2 Pro.

For anyone shipping a Mac or iOS app with transcription built in, that's the whole pitch. It runs on-device: better accuracy, better speed, no API calls, no server costs, no audio leaving the machine. That means one less external dependency, and one less privacy question about where audio goes.

Yes, but Whisper still covers far more languages and runs on Windows, Linux, and Android, where SpeechAnalyzer doesn't exist. The benchmark only covered clean audio, so it doesn't say how either model handles accents or background noise.

What I'd do

For an Apple-only app, I'd swap in SpeechAnalyzer for the English path. Keep Whisper around only if you need other languages or Windows and Android support. Running both works too: route by platform and language, fall back to Whisper where SpeechAnalyzer doesn't reach. What I wouldn't do is keep defaulting to Whisper out of habit on an Apple-only product.