Versions
What changed in each version of Seal Speak, and the story behind the bigger choices. Each version says honestly where it stands: only what is on the App Store is live.
Version 1.0
1.0.0 On the App Store
The first version, released on 28 October 2026. Get it on the App Store.
- Live captions of what people around you say, in large clear text, on your iPhone. It also runs on iPad, in the iPhone-sized layout; 1.0.1 brings the full-screen iPad layout.
- One big button to start and stop.
- A new line, labelled Speaker 1 or Speaker 2, when the voice changes. It goes by the sound of the voice, so it can mix people up.
- Text size that follows your phone, with A− and A+ to go bigger or smaller.
- Six colour themes, and six letterings including OpenDyslexic and Atkinson Hyperlegible.
- Captions scroll on their own. Scroll back to reread, then tap to jump to the newest line.
- Short labels for sounds like laughter, applause or a doorbell.
- Highlight words to copy them.
- Tap [ Save ] to keep a conversation on your phone for 30 days. Share it as text, or delete it sooner.
- Private: captions are made on your phone, audio is never recorded or sent anywhere, no account, no ads, no tracking.
- Works with no internet after a one-time setup.
- $4.99, once. English captions. iPhone 12 or newer with iOS 26 or later, or an iPad.
Captions are made by a computer and can contain mistakes. Seal Speak is not a medical device. The captions in 1.0.0 come from Apple's on-device speech recognition.
1.0.1 Coming next
Display fixes.
- A full-screen iPad layout, with the captions using the whole screen.
- It also works on iPhone Duo, folded or open.
Not on the App Store yet. No date.
Version 1.1
1.1.0 Planned
- A new engine makes the captions: Parakeet Ultra instead of Apple's speech recognition. In our tests it makes fewer mistakes and the words show up sooner. The story is below.
- Captions in more languages. Parakeet understands 25 European languages: Bulgarian, Croatian, Czech, Danish, Dutch, English, Estonian, Finnish, French, German, Greek, Hungarian, Italian, Latvian, Lithuanian, Maltese, Polish, Portuguese, Romanian, Russian, Slovak, Slovenian, Spanish, Swedish and Ukrainian. So far we have measured English, Spanish, French and German (below), and we test each of the others before switching it on.
- Live translation of the captions into another language, right on your phone. It is an optional add-on: $1.99, once, as an in-app purchase. Captions in every language stay part of the app. It uses Apple's on-device translation, the most accurate engine we tested (below).
- It still works with no internet, and nothing leaves your device.
- The app gets bigger: the new caption model is about 498 MB.
Planned, not built yet. The new engine has run on a Mac; we still need to check it on an iPhone. No date.
How we chose the caption engine
The engine is the part that turns speech into words. Everything else in Seal Speak depends on it, so we tested the candidates instead of guessing.
20%fewer mistakes than Apple's engine in English conversations: 13.8 in 100 words against 17.2
0.25 smedian wait for a correct word to show, against 0.67 s for Apple
Seal Speak 1.0 uses Apple's speech recognition, because it was the one built to run fully on the phone. Before settling on it for good, we wanted to know if something better exists, especially for people who speak with an accent. A caption app that works for one kind of voice is not good enough.
So we found a set of real conversations, checked by people, from speakers all over the world, and ran every engine on exactly the same recordings.
How we tested
We count a mistake the standard way: word error rate, the share of words that were wrong, missing or added. Lower is better. We aim for 15% or less (good) and treat 30% or less as acceptable. Each score comes with a 95% confidence interval, a range that shows how sure we are. When two ranges overlap a lot, the engines are close and we should not pretend one clearly wins.
Mistakes per 100 words, by engine
Show the numbers as a table
| Engine | WER | 95% CI | Correct word, median | Speed |
|---|---|---|---|---|
| Parakeet Ultra 6-bit | 13.8% | 12.8% to 14.8% | 0.25 s | 101.6x real time |
| Parakeet v2 | 15.5% | 14.5% to 16.6% | 0.24 s | 117.3x |
| sherpa-onnx | 16.1% | 14.9% to 17.2% | 0.39 s | 31.0x |
| Apple SpeechAnalyzer | 17.2% | 16.0% to 18.4% | 0.67 s | 43.6x |
Parakeet v2 and Parakeet Ultra made fewer mistakes than Apple's recognition, and Ultra was the most accurate. It was also quicker to show a correct word: a median of 0.25 seconds after it was said, against 0.67 seconds for Apple.
Accuracy was not the only question. Parakeet Ultra is a big model, and a big download is a cost for every user. So we tried making it smaller, and checked what that does to accuracy.
Size against mistakes, Parakeet builds
Show the numbers as a table
| Build | On disk | WER | 95% CI | Beats Apple? |
|---|---|---|---|---|
| Redux (2-bit) | 220 MB | 16.7% | 15.5% to 18.0% | no |
| v3 (4-bit) | 336 MB | 22.7% | 20.8% to 24.6% | no |
| Ultra 4-bit | 350 MB | 14.8% | 13.7% to 15.8% | yes |
| v2 | 464 MB | 15.5% | 14.5% to 16.6% | yes |
| Ultra 6-bit (chosen) | 498 MB | 13.8% | 12.8% to 14.8% | yes |
| Ultra 8-bit (as published) | 632 MB | 13.8% | 12.7% to 14.7% | yes |
The 6-bit version is smaller than the published 8-bit build (498 MB instead of 632 MB) and made no more mistakes in our test. The 4-bit version is smaller still (350 MB) and cost about one point of accuracy, though it still beat Apple. The ready-made small builds from others were either less accurate than Apple (Redux) or much worse (v3 4-bit).
What we chose, and why
Parakeet Ultra, 6-bit. Decided on 22 October 2026.
- It is tied for the fewest mistakes of anything we tested: 13.8 in 100 words, against 17.2 for Apple.
- It is 134 MB smaller than the 8-bit version and just as accurate (also 13.8%).
- It is fast to show words (median 0.25 s), and it runs on the phone with no internet.
If it does not work out on an iPhone, the fallback is Parakeet Ultra 4-bit: smaller (350 MB), about a point less accurate, still better than Apple.
What we have not shown yet: these tests ran on a Mac with one set of recordings. The next step is checking on an iPhone. Until then, Seal Speak uses Apple's recognition. The switch arrives in 1.1.0. Captions can still contain mistakes with any engine.
The full results, including every accent we tested, are in the engine accuracy report on GitHub. The research note behind the size test is evidence log E12.
How Parakeet did in other languages
Before promising a language, we measured it.
40%fewer mistakes than Apple in French, 39% in German and 30% in Spanish
We ran Apple's speech recognition and Parakeet Ultra on the same recordings in Spanish, French and German, and counted mistakes the same way as in the English test: word error rate, with a 95% confidence interval. Lower is better.
| Language | Apple | Parakeet Ultra |
|---|---|---|
| Spanish | 4.0% (3.4% to 4.5%) | 2.8% (2.4% to 3.3%) |
| French | 7.5% (6.7% to 8.3%) | 4.5% (3.9% to 5.1%) |
| German | 7.2% (6.5% to 7.9%) | 4.4% (3.8% to 5.0%) |
Parakeet made 30 to 40% fewer mistakes in every language, and the ranges do not overlap. That is why 1.1.0 uses it.
These recordings are people reading sentences in quiet rooms, not conversations, so every engine scores much better here than in real life. The ranking is the useful part. Measured on a Mac, not an iPhone. Source: language accuracy report (FLEURS test sets, CC BY 4.0).
How we chose the translation engine
Three engines can translate on the phone with no internet. We tested all three on the same sentences, then tested the winner in every language it offers.
19languages tested with Apple's translation, 1,012 sentences each
18 of 19scored as well as Spanish, French and German, where Apple beat both other engines
70 msor less to translate a caption line (median, on a Mac)
We score translations with chrF: how closely a translation matches one written by a person, from 0 to 100. Higher is better. We tested two kinds of text: ordinary written sentences, and the same sentences the way captions arrive, in lower case with no punctuation.
| English into | Text | Apple | OPUS-MT | Google ML Kit |
|---|---|---|---|---|
| Spanish | written | 56.8 | 55.0 | 53.2 |
| Spanish | caption-style | 54.8 | 51.2 | 51.8 |
| French | written | 73.2 | 69.8 | 65.4 |
| French | caption-style | 69.7 | 63.0 | 62.6 |
| German | written | 68.2 | 63.6 | 59.0 |
| German | caption-style | 65.9 | 58.2 | 56.7 |
Apple's translation scored highest everywhere, and it lost the least when the punctuation was missing, which is how captions arrive. All three were fast enough for live captions. That is why 1.1.0 uses Apple's translation.
Every language Apple translates into
Then we ran Apple's translation in all 19 languages it offered on our Mac, on the same 1,012 sentences. Here we add a second score, COMET, from 0 to 1, which judges whether the meaning came across and can be compared between languages. chrF can't: it always runs low for Japanese, Chinese and Korean, which don't put spaces between words.
| English into | chrF | COMET | Time per line |
|---|---|---|---|
| Indonesian | 73.8 | 0.920 | 53 ms |
| Portuguese (Brazil) | 73.6 | 0.899 | 51 ms |
| French | 73.2 | 0.879 | 58 ms |
| German | 68.2 | 0.883 | 70 ms |
| Turkish | 65.7 | 0.903 | 52 ms |
| Arabic | 63.1 | 0.870 | 53 ms |
| Vietnamese | 62.5 | 0.892 | 52 ms |
| Ukrainian | 61.8 | 0.898 | 51 ms |
| Hindi | 60.9 | 0.807 | 52 ms |
| Italian | 60.4 | 0.879 | 52 ms |
| Russian | 60.0 | 0.886 | 52 ms |
| Dutch | 59.6 | 0.879 | 51 ms |
| Thai | 57.9 | 0.879 | 51 ms |
| Spanish | 56.8 | 0.863 | 59 ms |
| Polish | 53.8 | 0.889 | 52 ms |
| Japanese | 43.5 | 0.912 | 59 ms |
| Chinese (Simplified) | 40.3 | 0.871 | 56 ms |
| Korean | 38.0 | 0.899 | 56 ms |
| Chinese (Traditional) | 32.6 | 0.875 | 55 ms |
- 18 of the 19 scored a COMET of 0.86 or more, the same range as Spanish, French and German.
- Hindi scored lowest (0.807). It needs a closer look before we offer it.
- Simplified Chinese sometimes came back in Traditional characters, in about 2 of every 100 caption lines. We will check this on an iPhone.
- Each language is a one-time download of 154 to 291 MB, made by iOS, not by Seal Speak.
Which of these languages 1.1.0 offers is not decided yet.
Sentences from FLORES-200 (1,012 per language), written text rather than conversation. Measured on a Mac, not an iPhone. Sources: translation comparison report, with every confidence interval, and results for every language.