Why I made Seal Speak
Everyone deserves to be part of the conversation.
I built Seal Speak as a birthday gift for my mom, who is hard of hearing. Now it's here for you too.
If it helps you stay close to the people you love, it's for you.
iPhone
Follow the conversation at the table.
Seal Speak shows what the people around you are saying, as they say it, in large clear text. It all happens on your iPhone, so it works with no internet.
Download on the App Store For iPhone 12 or newer with iOS 26 or later. English captions for now.- Made on your phone
- Works with no internet
- $4.99 once, no subscription
- No account
- No ads
ReadyNo one is talking right now
Speaker 1
Is it on? Say something so we can see.Speaker 2
Hi. My words are showing up on the screen.Speaker 1
It works in airplane mode too. Nothing leaves the phone.[Laughter]
Speaker 2
And it starts a new line when someone
Start a new conversation?
This one isn't saved.
Settings
How your captions will look
Speaker 1
Let's book a table for seven.Colors
Size
Lettering
OpenDyslexic was made for people with dyslexia. Atkinson Hyperlegible was designed by the Braille Institute for low vision.
Stop when it's quiet
How to use Seal
About
Saved
Kept for 30 days, then deleted.
Conversations you save show up here.
Delete this conversation?
This can't be undone.
This demo plays a made-up conversation. Tap Start captions, then ✕ to pause. Try [ Save ], tap the dim to read, open the list icon for saved conversations, and the gear for Settings.
What it does
One screen and one button to start. Nothing to set up each time.
Put your phone on the table
Open Seal Speak and tap Start captions. Words show up as people talk.
Read along
When the voice changes, Seal starts a new line and labels it Speaker 1 or Speaker 2. Sounds like laughter or a doorbell show up as short labels.
Pause when you need to
Tap ✕ and the words dim, with [ Save ] and [ New ] on top. Tap the dim to scroll back and reread. Hold an empty spot to bring the buttons back.
Copy a few words
Hold a word to select it, drag the handles to take in more, then tap [ Copy ]. Paste it into a message or a note.
Keep a conversation
Tap [ Save ]. Your saved conversations are behind the list icon at the top right. Each one stays on your phone for 30 days, then it's deleted.
Share it if you want
Open a saved conversation and tap [ Share ] to send the text to Notes, Messages, Mail, Files or AirDrop. [ Delete ] gets rid of it now.
Nothing leaves your device.
- Never recordedThe sound is turned into words, then gone.
- No internetWorks in airplane mode after a one-time setup.
- No accountNo ads, no tracking. Privacy policy
Made to be easy to read
Everything on screen fades in and out. Nothing slides around while you read.
Is it on? Say something so we can see.
Large text
Follows your phone's text size, with A− and A+ to go bigger or smaller.
Six color themes
Three light and three dark. Every one keeps strong contrast between the words and the background.
Six letterings
OpenDyslexic, Atkinson Hyperlegible, Standard, Rounded, Serif and Typewriter. Pick whichever is most comfortable.
Why Seal looks and works this way
The choices behind the design, and where they come from.
Contrast between light and dark matters most
Readers with low vision read better with strong brightness contrast than with color contrast alone. So every theme is built for brightness contrast, and a test makes sure each one reaches at least 7:1.
Source: Legge et al. (1990), Psychophysics of reading XI.
You choose the size, lettering and colors
Caption size, lettering, spacing and the contrast between caption and background all affect how easy captions are to read. Seal lets you change the size, the lettering and the colors.
Source: a caption readability study with older TV viewers, Sun Repository.
Light and dark, offered equally
Research found dark text on a light background read a little better, with a smaller difference for older adults, and dark mode caused no extra eye strain. Both work, so Seal offers three of each instead of choosing for you.
Source: Piepenbrock et al. (2013), summarised by the Nielsen Norman Group. The study was with people with healthy eyes.
OpenDyslexic, offered for comfort
Seal's default lettering is OpenDyslexic. Studies haven't shown it makes reading faster or more accurate, so it's offered because some people find it comfortable, alongside Atkinson Hyperlegible (designed by the Braille Institute for low vision) and four others.
Source: Wery & Diliberto (2017), Annals of Dyslexia.
Built on proven on-device tools
Speech-to-text uses Apple's on-device speech recognition. Telling voices apart uses the open-source FluidAudio library and NVIDIA's Sortformer model, bundled in the app so nothing goes over the network.
How we test Seal Speak
We don't guess which parts work best. We try each option on the same real recordings and count the mistakes.
31hours of real conversations, from 122 people with accents from all over the world
20%fewer mistakes in English with the caption engine coming in 1.1
0.25 sfor a word to show, against 0.67 s today (median, on a Mac)
Fewer mistakes in conversation
Seal Speak 1.0 uses Apple's speech recognition. We tested it against three other engines, and Parakeet Ultra made the fewest mistakes. It comes in 1.1.
| Mistakes in every 100 words | Today: Apple | In 1.1: Parakeet |
|---|---|---|
| English conversation | 17.2 | 13.8 |
Lower is better. 17,528 clips of people talking, the same clips for both.
Other languages too
Captions are English only today. Before offering more languages, we measured them. In Spanish, French and German, Parakeet made 30 to 40% fewer mistakes than Apple.
| Mistakes in every 100 words | Apple | Parakeet |
|---|---|---|
| Spanish | 4.0 | 2.8 |
| French | 7.5 | 4.5 |
| German | 7.2 | 4.4 |
These are people reading sentences aloud in a quiet room, so every engine does better here than in a real conversation.
For the curious: how we tested
Captions in English
- Recordings: the EdAcc corpus, 31.4 hours of conversation in many English accents, checked by people (CC BY-SA). 17,528 clips from 122 speakers.
- Every engine ran on exactly the same clips. A mistake is the standard word error rate: words that were wrong, missing or added, out of all the words said.
- Each score has a 95% confidence interval from a speaker-level bootstrap. When two ranges overlap, we don't call a winner.
- Timing: 526 clips (45 minutes) fed in as if live. "Correct word" is when a word first shows correctly, measured from when it was said.
| Engine | WER | 95% CI | Correct word, median |
|---|---|---|---|
| Parakeet Ultra 6-bit (1.1) | 13.8% | 12.8% to 14.8% | 0.25 s |
| Parakeet v2 | 15.5% | 14.5% to 16.6% | 0.24 s |
| sherpa-onnx | 16.1% | 14.9% to 17.2% | 0.39 s |
| Apple SpeechAnalyzer (1.0) | 17.2% | 16.0% to 18.4% | 0.67 s |
Other languages
- FLEURS test sets (CC BY 4.0): read speech, 908 Spanish, 676 French and German clips. Accent marks count: esta for está is a mistake.
- Spanish 4.0% (3.4 to 4.5) against 2.8% (2.4 to 3.3); French 7.5% (6.7 to 8.3) against 4.5% (3.9 to 5.1); German 7.2% (6.5 to 7.9) against 4.4% (3.8 to 5.0). The ranges don't overlap.
Measured on a Mac, not yet on an iPhone. Reports: captions in English, other languages.
Captions are made by a computer and can contain mistakes with any engine.
Questions
Which devices does it work on?
iPhone 12 or newer with iOS 26 or later, and iPad with an A14 chip or newer on iPadOS 26 or later. The Mac version is in testing.
Does it record or upload conversations?
No. Sound is turned into text on your device and the audio is never kept. Nothing is sent anywhere, and after a one-time setup it needs no internet at all. Text is kept only if you tap [ Save ].
How long are saved conversations kept?
30 days, then they're deleted on their own. They aren't backed up to iCloud. Open one from the list icon and tap [ Delete ] to remove it sooner.
What if I forget to stop it?
Seal stops by itself after 5 minutes with no one talking. In Settings you can change that to 15 or 30 minutes, or never.
More answers on the support page.
iPad
Talk across the table, in two languages.
Lay the iPad flat between you. Each of you reads the conversation in your own language.
- TodayRuns in the iPhone layout
- 1.0.1The whole screen
- PlannedTwo languages, split screen
A made-up conversation. The top half is upside down because it faces them.
Four times the screen.
Same words, same number of lines, letters about two and a half times taller. Readable from across a table.
19 languages, as fast as people talk.
70 msor less to translate a line
18of 19 languages as good as Spanish, French and German
- Indonesian0.920
- Japanese0.912
- Turkish0.903
- Portuguese0.899
- Korean0.899
- Ukrainian0.898
- Vietnamese0.892
- Polish0.889
- Russian0.886
- German0.883
- French0.879
- Italian0.879
- Dutch0.879
- Thai0.879
- Chinese (Trad.)0.875
- Chinese (Simpl.)0.871
- Arabic0.870
- Spanish0.863
- Hindi0.807
The line marks 0.86, where Spanish, French and German sit. Apple's translation beat two other engines in those three. Hindi needs a closer look; which languages 1.1 offers isn't decided. Written sentences, on a Mac. Every result.
How we tested translation
Three translators, the same 1,012 sentences. Then the same sentences the way captions arrive: lower case, no punctuation.
| Closeness to a person, 0 to 100 | Apple | OPUS-MT | Google ML Kit |
|---|---|---|---|
| Spanish | 56.8 | 55.0 | 53.2 |
| as captions | 54.8 | 51.2 | 51.8 |
| French | 73.2 | 69.8 | 65.4 |
| as captions | 69.7 | 63.0 | 62.6 |
| German | 68.2 | 63.6 | 59.0 |
| as captions | 65.9 | 58.2 | 56.7 |
Higher is better. Apple's lost the least when the punctuation was gone, which is how captions arrive. FLORES-200 sentences, on a Mac. The report.
Questions
Which iPads does it work on?
Any iPad with an A14 chip or newer, on iPadOS 26 or later.
Does translation need the internet?
Only once per language, to download it: 154 to 291 MB, from Apple. After that it translates with no connection, and nothing leaves the iPad.
When does the split screen come?
It's planned, with no date yet. Captions in more languages and translating the captions come first, in 1.1.
Can I use it on iPad today?
Yes. Seal Speak 1.0 runs on iPad in the iPhone-sized layout, with everything the iPhone app does.
More answers on the support page.
Mac
Captions for your calls.
A small window that stays on top of your call. ⌘H hides it.
- In testingNot on the App Store yet
- NeedsApple silicon (M1 or later), macOS 26
A made-up call.
Any call app.
Seal Speak captions the sound your Mac plays, so the app doesn't have to support it.
The room and the call, one conversation.
Everyone gets their own label, whether they're beside you or on the screen.
A call rings. Seal offers to caption it.
- You're getting a callCaption callNot now
It rings
- Captioning your FaceTime callHi! Can you read me okay?
It captions
- Call endedSave transcriptDon't save
You choose
FaceTime and iPhone calls. For other apps, press Start.
How we're testing calls
The Mac app is new. Here's what has been checked, and what hasn't yet.
1.0 sto switch from the room to a call that rings
- A call's voices, captioned word for word
- The room and the call at the same time
- A call rings, is captioned, then the room comes back
- ⌘H hides the window
- A real FaceTime call ringing
- Zoom, Teams and Google Meet on a real call
- Speakers turned up, for echo
- A full table of people in the room
So far the call was a stand-in: speech played through the Mac. Words said during the 1.0 s switch can be lost.
Questions
Does it record my calls?
No. The call's sound is turned into text on your Mac and isn't kept. The words are kept only if you choose [ Save transcript ] when the call ends.
Do the people on the call see it?
No. Seal Speak doesn't join the call. It only listens to the sound your Mac plays.
Does it work with headphones?
It's built to. With headphones that have a microphone, it captions you and the call; with headphones that don't, the call alone. That part is still being tested.
When can I get it?
It's in testing now, with no date yet. It needs a Mac with Apple silicon (M1 or later) on macOS 26 or later. It won't run on Intel Macs.
More answers on the support page.