Inscribe's own tests put SpeechAnalyzer ahead of every Whisper model it ships at roughly three times the speed — but a separate Argmax benchmark finds it loses to Apple's own older engine on keywords.
Apple's new SpeechAnalyzer, the on-device speech engine in iOS 26 and macOS 26, is the most accurate on-device English speech engine Inscribe tested, the company reported after benchmarking it against the Whisper models developers have long defaulted to.
On LibriSpeech read speech, Inscribe measured a 2.12% word error rate for SpeechAnalyzer on clean audio and 4.56% on the noisier test set, ahead of Whisper Small's 3.74% and 7.95% — while running about three times faster per second of audio.
The comparison mattered because Apple shipped its new SpeechAnalyzer and SpeechTranscriber APIs — the replacement for SFSpeechRecognizer — without publishing any accuracy figures, leaving developers to guess how they would perform.
For developers still on the old API, the generational jump is large. Inscribe found the new API cuts word error rate by roughly 3.5 to 4 times versus SFSpeechRecognizer on the same audio, with no accuracy trade-off in its measurements. The legacy engine came last on clean speech at a 9.02% word error rate, trailing even the roughly 40MB Whisper Tiny model.
Inscribe has a stake in the outcome: its on-device AI workspace ships both of Apple's engines alongside three Whisper models. To address that, it validated its harness by reproducing OpenAI's own published Whisper word-error rates on the same benchmark — its numbers matched closely, with a small consistent offset it attributed to normalization and quantization differences — and it published the raw per-utterance transcripts from both Apple engines across all 5,559 LibriSpeech utterances for outside rescoring.
The English win is bounded. Whisper models still cover a far wider spread of languages — over 100 in total, against roughly 30 locales for Apple's SpeechTranscriber — and Whisper can run on platforms beyond Apple devices on OS 26. Acting on its own results, Inscribe changed its app's automatic engine selection to prefer SpeechAnalyzer for supported languages and fall back to Whisper for the rest.
The new engine is not better at everything. In a separate Argmax benchmark on conversational earnings-call audio, the older SFSpeechRecognizer beat SpeechAnalyzer on known-keyword accuracy — a gap Argmax pins on Custom Vocabulary, the older API's feature for registering keywords that the new one drops.