Apple SpeechAnalyzer vs Whisper (2026): What the Headline Benchmark Left Out

Quick answer: In the benchmark everyone shared, Apple’s new on-device SpeechAnalyzer did beat Whisper — 2.12% word error rate against 3.74% on LibriSpeech test-clean. But that test ran Whisper Small, on English read audiobooks, on a Mac. It never tested Whisper Large V3 or Large V3 Turbo, the models an on-device iPhone app actually loads, and it never touched a meeting, an accent, or any of Whisper’s other ~99 languages. Apple’s engine is genuinely fast, genuinely accurate, and free — inside the 22 languages it supports. Whisper still covers about five times more.

Every number below comes from the benchmark author’s own write-up, OpenAI’s published papers and model cards, the Open ASR Leaderboard, or Apple’s developer documentation — all linked in the sources at the end. It was written by the LoroNote team on August 23, 2026. LoroNote runs Whisper Large V3 Turbo on the device, so we have a stake in this comparison. That is exactly why every figure here is someone else’s measurement, not ours.

What SpeechAnalyzer actually is

SpeechAnalyzer is the speech framework Apple introduced at WWDC 2025 and shipped in iOS 26. It replaces SFSpeechRecognizer, the aging API that powered most third-party dictation on iOS for a decade.

Two public modules matter:

• SpeechTranscriber does speech-to-text.
• SpeechDetector flags where speech is present in an audio stream.

Everything runs on device. Language assets download through the system asset catalog rather than being bundled by each app, which is why an app using it adds almost nothing to its download size. For locales SpeechTranscriber does not cover, Apple provides DictationTranscriber as a fallback.

The hard requirement is the one most comparisons skip: SpeechAnalyzer needs iOS 26 or later and is not backward compatible. On an iPhone still running iOS 18, it does not exist.

The benchmark everyone cited

The numbers that spread across tech media in July 2026 come from one developer-run evaluation. Here is what it actually measured:

And the results:

Two things in that table are worth saying plainly. Apple’s win over its own legacy API is enormous — roughly a 75% error reduction, and the real story for anyone who built on SFSpeechRecognizer. And Apple genuinely beat every Whisper model that was tested.

The problem is which models were tested.

Three things the benchmark did not measure

1. The Whisper models that on-device apps actually run

Whisper Small is a 460 MB model. It is not what a serious transcription app loads on a modern iPhone. LoroNote runs Large V3 Turbo — 809 million parameters, with the decoder cut from 32 layers to 4 for speed.

The benchmark stopped before Medium, Large, and Turbo. For scale, published evaluations put Whisper Large V3 on LibriSpeech at roughly 2.0–2.5% WER on test-clean and about 3.9–4.3% on test-other — the same neighbourhood as Apple’s 2.12% and 4.56%, or slightly better.

Treat that comparison carefully. Those Whisper figures come from separate evaluations with their own text normalization, not from a head-to-head run on the same machine against Apple’s engine. Normalization choices alone can move LibriSpeech WER by more than a point — our WER guide explains how that happens. The honest conclusion is not “Whisper Large wins” — it is that no published test has compared Apple’s engine against the Whisper models the comparison is actually about, and the gap the headlines implied is an artifact of the model size that was chosen.

2. Anything harder than a studio audiobook

LibriSpeech is read speech: one person, reading a book, into a good microphone, with no cross-talk. It is the easiest rung on the ladder, which is why every engine posts its best number there.

Real recordings sit further down. On the accuracy ladder we documented for Whisper, the same Large V3 model that scores ~2–3% on clean read speech averages about 7.4% across mixed real English benchmarks and around 16% on AMI meeting-room audio with distant microphones and overlapping speakers.

Nobody has published those rungs for SpeechAnalyzer. Until someone does, a 2.12% audiobook score tells you very little about a four-person meeting recorded from the middle of a table. The benchmark’s author was upfront about this — the write-up describes the corpus as “read audiobook speech, not meetings.”

3. Ninety-nine other languages

This is the largest gap, and the benchmark author flagged it directly: the numbers “say nothing about the 100+ languages Whisper supports.”

As of August 2026, SpeechTranscriber.supportedLocales returns 42 locales covering 22 languages:

Arabic, Chinese (Simplified, Traditional, Hong Kong), Cantonese, Danish, Dutch, English (nine regional variants), Finnish, French, German, Hebrew, Italian, Japanese, Korean, Malay, Norwegian Bokmål, Portuguese, Russian, Spanish, Swedish, Thai, Turkish, and Vietnamese.

Whisper Large V3 covers about 100. The absences in Apple’s list are not obscure: no Hindi, no Polish, no Czech, no Greek, no Ukrainian, no Indonesian. Portuguese ships as pt_BR only — there is no European Portuguese locale.

Apple has said more languages are coming, and this list will be out of date eventually. But “supported” and “equally accurate” are different claims for both engines, which is why the full language roster is worth checking against your own audio rather than trusting either count.

About that “3x faster” number

The speed claim travelled as widely as the accuracy one, and it has the same footnote: it was measured against Whisper Small.

What the benchmark reported was a throughput range of roughly 12x to 40x realtime on an M2 Pro — an hour of audio in about 1.5 to 5 minutes. The author deliberately withheld precise per-model timings pending cleaner measurements on an idle machine.

That range is real and Apple’s engine is fast. It also came from a desktop M-series chip with 32 GB of RAM, not from a phone. On iPhone, the binding constraint for any transcription model is memory rather than raw compute — which is why several well-known “Whisper apps” quietly load small or medium on the phone while their Mac versions run large. We listed which variant each iOS app actually loads if you want to check the one you use.

Where each one is the better choice

Neither engine is the general winner. They are built for different jobs.

Choose Apple SpeechAnalyzer when:

• Your language is in the 22 it supports, and especially if it is English.
• You are on iOS 26 or later.
• You want live, streaming transcription with partial results updating as someone speaks — this is what the API is designed around.
• You want zero model download and zero cost.

Choose the Whisper Large V3 family when:

• Your language is outside Apple’s list, or your recordings mix languages within one file.
• You need the same behaviour on older devices — Whisper-based apps run well below iOS 26.
• You are transcribing existing files rather than live speech, where accuracy matters more than latency.
• You want the transcript to come out the same regardless of which OS version the device is on.

There is no free lunch in either direction. Apple’s engine is faster to integrate, free, and excellent inside its range. Whisper’s range is roughly five times wider, and the Large variants carry that breadth at a real cost in model size and processing time.

What this means when you pick an iPhone app

Two practical consequences.

First, “AI transcription” on the App Store now means at least three different things: an app calling Apple’s on-device SpeechAnalyzer, an app running its own Whisper model on device, or an app uploading your audio to a server. Those have completely different privacy, offline, and language properties, and the listings rarely say which. Our comparison of iOS speech-to-text apps sorts them by where the audio is processed.

Second, an app being “on-device” no longer implies anything about accuracy. Both routes are on-device. The question shifted from where does my audio go to which model, at what size, in which language.

Test it yourself in five minutes

Benchmarks are averages over corpora you will never record. Your own audio is the only test that answers your question:

1. Record two minutes in the room you normally record in — your microphone, your accent, a few names, numbers, and technical terms.
2. Transcribe it with your iPhone’s built-in transcription (Voice Memos on a supported device, or any app using Apple’s engine).
3. Transcribe the same file with a Whisper-based app.
4. Count errors in a 200-word stretch: every substituted, missing, or invented word. Errors ÷ 200 = your WER. (The full method, including how to decide what counts as an error, is in our WER guide.)

If both come back clean, use whichever is more convenient. The difference tends to show up on proper nouns, overlapping speech, and any language outside the high-resource tier — which is exactly where the benchmark stopped looking.

Frequently asked questions

Is Apple SpeechAnalyzer more accurate than Whisper?

Against Whisper Small on clean English audiobooks, yes — 2.12% versus 3.74% WER. Against Whisper Large V3 or Large V3 Turbo, nobody has published a head-to-head test. Separate evaluations put Large V3 in the same range or slightly ahead on the same corpus, but different normalization makes that a rough comparison rather than a verdict.

Does SpeechAnalyzer work offline?

Yes. It runs entirely on device, with language assets downloaded once through the system asset catalog. In that respect it matches an on-device Whisper app — neither sends your audio to a server.

Which languages does SpeechAnalyzer support?

42 locales across 22 languages as of August 2026, including English, Spanish, French, German, Italian, Japanese, Korean, Chinese, Cantonese, Arabic, Russian, Portuguese (Brazil only), Dutch, Turkish, Thai, Vietnamese, Malay, Hebrew, Danish, Finnish, Norwegian, and Swedish. Whisper covers about 100 languages. Apple has said the list will grow.

Do I need iOS 26 to use it?

Yes. SpeechAnalyzer requires iOS 26 or later and is not backward compatible. Apps built on Whisper run on considerably older versions, which is one of the practical reasons to keep both engines in mind.

Does either one identify who is speaking?

Speaker identification is a separate step from transcription, and neither engine performs it as part of speech-to-text. Apps add it on top — how speaker separation works covers what that involves and where it breaks down.

Why does LoroNote use Whisper instead of Apple’s engine?

Language coverage and device range. LoroNote transcribes about 100 languages, including mixed-language recordings, on devices going back well before iOS 26. Apple’s engine is excellent within its 22 languages; it is not a drop-in replacement for that breadth. We explained the specific model choice in our Whisper Large V3 Turbo guide.

Sources

• Apple, Bring advanced speech-to-text to your app with SpeechAnalyzer — WWDC25 session introducing the framework
• Apple, SpeechTranscriber documentation
• Lyonesse, Apple’s New Speech API vs Whisper: The First Real Benchmark — the LibriSpeech evaluation, its methodology and its stated caveats
• OpenAI, Robust Speech Recognition via Large-Scale Weak Supervision — the Whisper paper
• OpenAI, whisper-large-v3 model card and whisper-large-v3-turbo model card
• Open ASR Leaderboard — mixed-benchmark and AMI meeting figures

Benchmark figures verified on August 23, 2026. Apple’s supported-locale list changes with OS updates; check SpeechTranscriber.supportedLocales on a current build before relying on it.