Published: August 2, 2026
Apple’s new SpeechAnalyzer speech-to-text API beat OpenAI’s Whisper Small on accuracy and speed in the first independent benchmark of the engine, published July 13, 2026 by transcription-app maker Inscribe. SpeechAnalyzer scored a 2.12% word error rate on clean speech versus Whisper Small’s 3.74%, while transcribing roughly 3x faster, across 5,559 LibriSpeech utterances on an Apple M2 Pro. Apple shipped the API with iOS 26 and macOS 26 without publishing any accuracy figures, so Inscribe’s test is the first public measurement of the engine that replaced the decade-old SFSpeechRecognizer. The benchmark reached 509 points on Hacker News, where it was the top story on July 14.
How SpeechAnalyzer scored against Whisper
SpeechAnalyzer beat all 3 Whisper models tested — Small, Base, and Tiny — on both the clean and noisy halves of the LibriSpeech corpus. The test covered 2,620 clean utterances (test-clean) and 2,939 noisier utterances (test-other), with every engine running fully offline:
| Engine | Clean speech WER | Noisy speech WER | Model size |
|---|---|---|---|
| Apple SpeechAnalyzer | 2.12% | 4.56% | built into OS 26 |
| Whisper Small | 3.74% | 7.95% | ~460MB |
| Whisper Base | 5.42% | 12.51% | ~140MB |
| Whisper Tiny | 7.88% | 17.04% | ~40MB |
| Apple SFSpeechRecognizer (legacy) | 9.02% | 16.25% | built-in |
The most lopsided result targets Apple’s own legacy engine. SFSpeechRecognizer, the API SpeechAnalyzer replaces, finished last on clean speech — behind even Whisper Tiny, a 40MB model. According to Inscribe, the new API cuts word error rate “about 3.5 to 4x”: an hour-long meeting transcribed with the legacy engine contains roughly 4 times as many wrong words as the same meeting through SpeechAnalyzer. All 5 engines ran faster than real time, transcribing an hour of audio in roughly 1.5 to 5 minutes on-device.
Why the numbers are credible despite the vendor’s stake
Inscribe sells a transcription app built on these engines, and the company designed the benchmark to survive that suspicion. Its Whisper measurements reproduce OpenAI’s own published LibriSpeech scores within 0.11 to 0.42 percentage points on all 6 checkable measurements, which validates the corpus handling, text normalization, and scoring that also produced the Apple numbers. Inscribe released the raw per-utterance transcripts for both Apple engines — all 5,559 utterances — so anyone can rescore the results. A follow-up round published July 15 added NVIDIA Parakeet and MOSS-Transcribe-Diarize on the identical corpus.
The limits: 30 locales, Apple Silicon only, no fine-tuning
SpeechAnalyzer carries 3 restrictions that keep Whisper relevant. The API supports roughly 30 locales against Whisper’s 100+ languages, runs only on Apple Silicon devices with OS 26 or later, and is a closed system model that cannot be fine-tuned on domain vocabulary, such as medical, legal, or product terminology. The benchmark also measured English read speech only — LibriSpeech contains audiobook recordings, not accented, far-field, or multi-speaker meeting audio. For cross-platform products or multilingual transcription, Whisper remains the realistic on-device choice.
What this means for businesses choosing transcription tools
Companies transcribing sensitive audio, such as client calls, medical dictation, or internal meetings, now have an on-device English option that outperforms the open-source standard on Apple hardware. On-device transcription keeps audio off third-party servers entirely, which simplifies compliance reviews that cloud APIs from Google, Microsoft, and OpenAI trigger. Buyers should expect Mac and iPhone transcription apps to migrate to SpeechAnalyzer over the next release cycles — Inscribe switched its own default engine after running the test. Our guide to the 15 best AI tools for business in 2026 covers transcription alongside the other categories where on-device options are displacing cloud subscriptions.
Our Take
The headline is Apple beating Whisper, but the business story is Apple beating Apple: the built-in speech engine every Mac and iPhone app could call for free was worse than a 40MB open-source model, and now it leads the field. When the OS ships a best-in-class engine at zero marginal cost, paid transcription tools have to justify themselves on workflow, not accuracy. Ask your transcription vendor which engine they run — and what you are paying for beyond it.
For Context
WithO2 tracks Apple’s AI positioning as it hardens into products: Apple’s lawsuit accusing OpenAI of stealing hardware secrets frames the rivalry behind this benchmark, and Apple’s $30 billion Broadcom chip deal covers the silicon investment that makes on-device AI viable.
Related reading: voice AI is consolidating across platforms — see ChatGPT Voice arriving on desktop and Claude’s voice mode automating Gmail, Slack, and Canva.
