Since macOS 26 and iOS 26, Apple has had a new on-device speech recognition API — SpeechAnalyzer and SpeechTranscriber, a new generation of its own recognition engine. We shorten it to Apple Speech below. Whisper Notes carries Whisper, Parakeet, SenseVoice and Qwen3-ASR side by side, and every one of them is there for a reason: some are extremely fast, some are more accurate in a particular language. Which makes the question worth asking properly: is Apple's built-in speech recognition good enough now? And how far is it from these on-device open models? We ran a strict test of our own. Here are the results.
How far is Apple Speech from the models you download?
| Language | Apple Speech | Qwen3-ASR | Whisper Large | Whisper Small | Parakeet | SenseVoice |
|---|---|---|---|---|---|---|
| English | 8.80 | 4.49 | 5.49 | 7.04 | 6.89 | 8.46 |
| German | 6.26 | 2.85 | 4.05 | 8.67 | 6.26 | — |
| Dutch | — | 7.36 | 5.69 | 16.75 | 8.58 | — |
| Russian | — | 5.65 | 5.13 | 11.37 | 7.12 | — |
| Polish | — | 12.59 | 5.45 | 15.81 | 8.30 | — |
| Japanese | 6.36 | 5.74 | 5.30 | 11.37 | — | 6.78 |
| Korean | 4.31 | 3.54 | 4.25 | 7.30 | — | 7.85 |
| French | 7.61 | 4.57 | 5.97 | 13.24 | 5.83 | — |
| Chinese | 7.97 | 6.49 | 7.97 | 20.50 | — | 7.69 |
| Spanish (Latin America) | 5.41 | 3.38 | 3.62 | 6.22 | 4.86 | — |
Ten of the most-used languages, one engine per column. Each cell is that engine's error rate, lower is better; green marks the lowest in the row, white the second-lowest. CER for Japanese, Korean and Chinese, WER for the rest, never merged into one number. Our own FLEURS run, one M5 MacBook Air, read speech.
Short names: Qwen3-ASR = Qwen3-ASR 1.7B, Whisper Large = Whisper Large V3 Turbo, Parakeet = Parakeet V3, SenseVoice = SenseVoice Small. A dash means the engine has no model for that language.
Two engines take all ten first places. Qwen3-ASR 1.7B wins six rows, Whisper Large V3 Turbo the other four. Apple Speech wins none. But on Korean, Japanese and Chinese it finishes within a point and a half of the leader, and on Chinese it lands exactly level with Whisper Large V3 Turbo at 7.97. Spanish and French are further behind, German further still. English is the floor: 8.80 against 4.49, last of six, beaten by every model you could have downloaded instead, including the smallest one. Three rows have no Apple cell at all; that is the next section. The honest summary has two halves: good enough to use where it is strong, and still a clear step behind the open models on every language we could compare.
Which languages does Apple Speech support?
More than twenty: in our run Apple Speech produced results in 21 of the 83 language configurations we tested, against 83 for both Whisper builds, 30 for Qwen3-ASR 1.7B and 25 for Parakeet V3. Of the ten languages above it is missing three — Dutch, Polish and Russian — where the API answers that it has no model, so there is nothing to benchmark. And there is no fixed list to quote: the language assets belong to macOS, so an app has to ask at runtime what a given machine actually has.
| Language | Apple Speech | Qwen3-ASR | Whisper Large | Whisper Small | Parakeet | SenseVoice |
|---|---|---|---|---|---|---|
| Italian | 4.39 | 2.58 | 2.58 | 7.64 | 3.43 | — |
| Cantonese | 8.69 | 6.44 | 36.75 | 119.01 | — | 37.23 |
| Portuguese (Brazil) | 9.35 | 5.17 | 5.44 | 8.61 | 6.49 | — |
The other languages Apple Speech measured cleanly here, ordered by its own error rate. Same columns, colours and short names as above; CER for Cantonese, WER for the other two. An error rate can pass 100% because insertions count in the numerator too.
Eleven further configurations it ran are left out — Bengali, Gujarati, Hindi, Kannada, Malayalam, Marathi, Nepali, Punjabi, Tamil, Telugu and Urdu — because Apple Speech answered in Latin transliteration instead of the native script, so the error rate there measures the script and not the recognition.
How we measured this
Fairness first. Every engine hears the same clips, in the same order, one at a time, on one machine: an M5 MacBook Air with 32 GB running macOS 27.0, so Apple's side is its current engine. The audio is the Google FLEURS test split at revision 70bb2e84, about 150 sentences and 30 minutes per language. We did not choose the sentences: a fixed hash of the dataset's own item IDs puts them in order, and we take the shortest whole-item run that crosses thirty minutes. No model output had any part in that choice. Every cell was rebuilt from its own receipt and checked again before we kept it, and all 285 passed.
Scores are word error rate where words are separated by spaces, and character error rate for Japanese, Korean, Chinese and Cantonese; the two are never merged into one number. One caveat outweighs the rest: FLEURS is read speech, not a meeting and not dictation. These tables tell you whether an engine is plausible for your language. They do not predict what you will get on your own recordings.
How much better than the old Apple Dictation?
Apple ships two transcription operating points, and we measured both. SpeechTranscriber is the new one, the one this post calls Apple Speech. DictationTranscriber is the compatibility one, the dictation path a Mac had before it.
| Language | Apple Dictation | Apple Speech |
|---|---|---|
| Korean | 7.11 | 4.31 |
| Italian | 6.93 | 4.39 |
| Spanish (Latin America) | 8.43 | 5.41 |
| German | 12.69 | 6.26 |
| Japanese | 9.37 | 6.36 |
| French | 17.10 | 7.61 |
| Chinese | 10.28 | 7.97 |
| Cantonese | 10.18 | 8.69 |
| English | 13.83 | 8.80 |
| Portuguese (Brazil) | 10.85 | 9.35 |
Every language both Apple operating points measured cleanly, ordered by Apple Speech's error rate; green marks the better value in the row. Same run, same clips, same Mac. CER for Japanese, Korean, Chinese and Cantonese, WER for the rest.
Apple Speech is more accurate on all ten. The median language loses about a third of its errors; French and German lose more than half. The smallest gains are Brazilian Portuguese and Cantonese, where roughly one error in seven goes away. If your sense of Apple's recognition comes from what dictation used to give you, that sense is out of date.
That is a comparison with Apple's own past, though, not with the field. Against the models you download, Apple Speech leads no row anywhere in this post; its best placing is second, on Cantonese. Dictation does still cover more languages — 33 configurations against 21 in our run, including all three the new engine is missing here.
What does the built-in engine buy you?
| Engine | Speed | Peak memory | Download |
|---|---|---|---|
| SenseVoice Small | 162.3x realtime | 0.95 GiB | 827 MB |
| Parakeet V3 | 75.6x realtime | 0.09 GiB | 465 MB |
| Apple Speech | 30.8x realtime | not reported | none, part of the system |
| Qwen3-ASR 1.7B | 11.1x realtime | 2.43 GiB | about 2.5 GB |
| Whisper Small | 7.9x realtime | 0.87 GiB | 600 MB |
| Whisper Large V3 Turbo | 3.0x realtime | 1.87 GiB | about 1.6 GB |
Speed is each engine's median across the languages in this post that it ran — isolated-item processing throughput, a diagnostic, not product latency. Peak memory is the process-group peak in this run. Apple Speech decodes inside a macOS service the calling app does not own, so no figure there would mean the same thing as the other five.
Apple Speech costs nothing to install and nothing to keep. There is no model file, no download, and nothing allocated in the app's own process. That last part is not the same as free: the system service does use memory while it works, and we cannot attribute it exactly, which is why we report no number there rather than a zero. It is also quick — about thirty times realtime here, behind only Parakeet V3 and SenseVoice Small. For a lot of people that is the whole argument, and on the languages where it is strong the accuracy is good enough to let the argument stand.
The engines that beat it charge for it. Whisper Large V3 Turbo wins four of the ten rows and is the slowest thing in this test, roughly ten times slower than Apple Speech, for about 1.6 GB on disk. Qwen3-ASR 1.7B wins the other six and asks for about 2.5 GB and 2.43 GiB of memory. Parakeet V3 is the counter-example at 465 MB and 0.09 GiB, and it pays unevenly: ahead of Turbo on French, well behind it on Polish, and it has no Japanese, Korean, Chinese or Cantonese at all. The closest comparison is Whisper Small, and it is the one that flatters Apple most — across the ten languages we measured cleanly for both, Apple Speech was more accurate on eight, and Whisper Small is a 600 MB file you have to keep.
Should you use Apple Speech instead of a downloaded model?
Language by language, this is how we would answer that.
- If you transcribe Dutch, Polish or Russian: Apple Speech is not an option, it has no model for them. Whisper Large V3 Turbo led all three.
- If disk space or download size is the constraint: Apple Speech installs nothing at all. Across the ten languages we measured cleanly for both, it beat Whisper Small on eight; English and Brazilian Portuguese were the exceptions.
- If you transcribe Korean, Japanese or Chinese and want the built-in engine: it finishes within a point and a half of the best model in the row on all three, and on Chinese it is level with Whisper Large V3 Turbo. That holds until you need the last point or two of accuracy.
- If you transcribe English: it was last of the six, 8.80 against 4.49 for Qwen3-ASR 1.7B and 5.49 for Whisper Large V3 Turbo. Download one of those.
- If you transcribe Cantonese: Apple Speech is second at 8.69, behind Qwen3-ASR 1.7B at 6.44 — and this is the one language where both Whisper builds are unusable, at 36.75 and 119.01.
- If you want the most accurate result per language and do not mind the download: Qwen3-ASR 1.7B led English, German, Korean, French, Chinese and Spanish; Whisper Large V3 Turbo led Dutch, Russian, Polish and Japanese.
- If you are on iPhone or the Mac App Store build: the engines Whisper Notes offers there are Whisper, Parakeet V3 and SenseVoice.
Apple Speech is our shorthand in this post for Apple's own SpeechTranscriber, the on-device API introduced in macOS 26 and iOS 26 that any Mac app can call. It is the system's engine rather than one of ours: there is no model file on disk, the language assets are managed by macOS, and it needs macOS 26 or later. It is not one of the models Whisper Notes for Mac downloads. Those model families are Whisper, Parakeet V3, SenseVoice and Qwen3-ASR on the Mac Direct Download (DMG) build, and Whisper, Parakeet V3 and SenseVoice on iPhone and in the Mac App Store build. The per-engine language table for those is on our language support page.
Nothing here changes what Whisper Notes for Mac does today: Parakeet V3 is still the default.
Frequently Asked Questions
Is Apple Speech as accurate as Whisper?
Not on most of the languages both cover, but the gap is smaller than the reputation. On our own FLEURS run, across the seven of these ten languages Apple Speech supports, Whisper Large V3 Turbo was more accurate on six and the two matched exactly on Chinese, 7.97% CER each. Against the smaller Whisper Small the result flips: Apple Speech was ahead on six of those seven, losing only English, 8.80% against 7.04% WER. Apple Speech leads no row in this test. It comes closest on Korean (4.31% CER against 3.54% for the row leader), Japanese (6.36% against 5.30%) and Chinese (7.97% against 6.49%), and it is furthest away on English, 8.80% against 4.49% WER. Cantonese is the one language we measured cleanly where Apple Speech beat Whisper Large V3 Turbo, 8.69% against 36.75% CER. The built-in engine makes sense when you want no download at all and your language is one it covers; Whisper Large V3 Turbo is the one to reach for when you want the most accurate result, or when your language is one of the three Apple Speech is missing here.
What languages does Apple Speech support?
More than twenty, though macOS decides the list rather than any app shipping one. In our run Apple Speech produced results in 21 of the 83 language configurations we tested, against 83 for both Whisper builds, 30 for Qwen3-ASR 1.7B and 25 for Parakeet V3. The main table in this post takes ten of the most-used languages so every engine can be compared on the same rows: there it has English, German, Japanese, Korean, French, Chinese and Spanish, and it does not have Dutch, Polish or Russian. Italian, Brazilian Portuguese and Cantonese are covered too, and their rows are in the second table. There is no fixed list to publish: an app has to ask macOS at runtime which languages a given machine has, because the assets are downloaded and managed by macOS, not by the app. If your language is missing, Whisper reaches every language in the app's list on every channel.
Apple Speech or Whisper Small if disk space is tight?
Apple Speech, on eight of the ten languages we measured cleanly for both. It scored 4.31% CER on Korean against Whisper Small's 7.30%, 4.39% against 7.64% WER on Italian, 5.41% against 6.22% on Spanish, 6.26% against 8.67% on German, 6.36% against 11.37% CER on Japanese, 7.61% against 13.24% on French, 7.97% against 20.50% CER on Chinese and 8.69% against 119.01% CER on Cantonese; Whisper Small was ahead on English, 7.04% against 8.80%, and on Brazilian Portuguese, 8.61% against 9.35%. Apple Speech also downloads nothing, while Whisper Small is a 600 MB model file. The reason to keep Whisper Small anyway is breadth: it covers every language in the app's list, including all three that Apple Speech is missing here, and it runs on every channel — iPhone, the Mac App Store build and the Direct Download build alike.
What is SpeechAnalyzer, and is it the same as Apple Dictation?
SpeechAnalyzer is the umbrella API Apple introduced at WWDC 2025 and shipped in macOS 26 and iOS 26. SpeechTranscriber is the transcription operating point inside it, and that is what we measured and what this post calls Apple Speech. Dictation is the other operating point, the compatibility one, and we ran it too: Apple Speech was more accurate on all ten languages both handled cleanly, taking about a third of the errors out of the median language and more than half out of French and German. The per-language table is in the section above. Dictation does cover more languages — 33 configurations against 21 in our run — so it is the system engine to reach for if you need Dutch, Polish or Russian and can accept 17.34%, 13.50% and 13.13% WER, where Whisper Large V3 Turbo scored 5.69%, 5.45% and 5.13% on the same clips.
Does audio leave my Mac when I use Apple Speech?
Apple documents SpeechTranscriber as on-device speech recognition, and that is the API we benchmarked here; the language assets are downloaded once by macOS and the recognition itself runs locally. What we measured in this post is accuracy and speed, not network behaviour, so we describe that part as Apple describes it. The engines Whisper Notes downloads itself — Whisper, Parakeet V3, SenseVoice and Qwen3-ASR — run on your own Mac's GPU or Neural Engine with no account and no API key, and transcription works with the network off, on every channel.
Why do these numbers not match the accuracy I get on my recordings?
Because FLEURS is short, clean, read speech, and your recordings are not. Our run is a calibration on a public dataset: about 150 sentences and 30 minutes of audio per language, one M5 MacBook Air, the same clips for every engine. It is useful for asking whether an engine is plausible for a language, and it is not a prediction of your accuracy on a meeting, on noisy audio, on accented speech or on a two-hour file.
Try it
The five downloadable models in this comparison — Whisper Small, Whisper Large V3 Turbo, Parakeet V3, SenseVoice Small and Qwen3-ASR 1.7B — are all in Whisper Notes for Mac, Direct Download (DMG). Settings, then Transcription Model. Read speech is a starting point; your own audio is the real test.
If you find a language where our numbers do not match your experience, email mac@whispernotes.app. Full release notes: whispernotes.app/changelog.
Sources for everything above: our FLEURS run on the Google FLEURS test split at revision 70bb2e84, Apple's SpeechAnalyzer documentation, and the thirty-language Qwen3-ASR run that came before it.
Download for macOS