For the past few years, if you needed on-device speech recognition and weren’t locked into Apple’s ecosystem, Whisper was the obvious default. OpenAI open-sourced it in 2022, the community built fast inference wrappers around it, and it covered 100 languages well enough to be a credible universal option. The question for Apple-platform developers was mostly whether to use Apple’s SFSpeechRecognizer (fast, private, limited) or Whisper (more capable, more flexible, heavier).

That calculus changed with Apple’s SpeechAnalyzer API, introduced in macOS 26.5.1. Inscribe, a transcription app, ran a straightforward benchmark: 5,559 utterances from the LibriSpeech clean-speech test set, word error rate measured across all models, raw transcripts published for verification.

The numbers are clear. SpeechAnalyzer achieves 2.12% WER on that corpus. Whisper Small comes in at 3.74%. Whisper Base at 5.42%, Tiny at 7.88%, and the legacy SFSpeechRecognizer at 9.02%. SpeechAnalyzer also runs roughly 3× faster than Whisper Small.

That’s not a marginal improvement. The old story was that Apple’s speech API was acceptable for simple voice commands but that Whisper was the better tool if accuracy mattered. At 2.12% vs 3.74% WER on clean speech, there’s no longer a meaningful accuracy argument for reaching for Whisper Small on Apple hardware.

The catch is language coverage. SpeechAnalyzer supports approximately 30 languages; Whisper supports 100+. If you’re building anything that needs multilingual support — which is increasingly common, even for apps ostensibly targeting a single market — Whisper still wins by default. Apple hasn’t announced a roadmap for expanding SpeechAnalyzer’s language support, and its release pattern suggests incremental additions rather than a sudden jump to 100 languages.

There’s also the usual Apple caveat: this API is Apple-only. Whisper can run on any OS, any hardware, any deployment target. SpeechAnalyzer only runs on Apple silicon through Apple’s frameworks. For teams shipping cross-platform or targeting server-side inference, this benchmark changes nothing.

But for teams that are already in the Apple ecosystem — iOS and macOS apps where user privacy and latency matter — the message is direct. You can now get better-than-Whisper-Small accuracy, faster, without shipping a model bundle of your own. The Neural Engine is doing the work. The benchmark methodology is public and reproducible, which is more than can be said for most vendor accuracy claims.

The broader pattern here is worth noting. Apple has been consistently competitive on per-inference quality for on-device tasks — face recognition, OCR, object detection — while staying quiet about it relative to the cloud model races. SpeechAnalyzer continues that pattern: quietly shipping something that matches or exceeds the open-source state of the art, scoped to Apple hardware, no blog post from Cupertino. The developers who notice are the ones who benchmark.