Why Caddie is the most accurate dictation in the real world

Every dictation app can type "the meeting is at three". That's not where they let you down.

They let you down on your own words. A colleague's name with an unusual spelling. The neighbourhood your office is in. The tools you say twenty times a day: Datadog, Convex, Next.js. And if you work in India, the sentence that's half Hindi: "kal tak bhej dena, it's for the Datadog review."

When an app gets one of those wrong, you stop. You find the word, select it, retype it. That takes about two seconds a word. The slowest app in our test waits about one second after you let go of the key. So one wrong word costs more than the whole wait.

The fix isn't a faster engine. It's an engine that already knows your words.

Dictation that lives inside your notes

A stand-alone dictation app meets you as a stranger. It has never seen your calendar or read a line you've written. So it learns slowly: you correct a word, or you type it into a dictionary by hand, and next time it might get it right. Wispr Flow works this way, and so do most others.

Caddie is where you already keep your notes, your journal, your meeting notes and your calendar. Those are full of the names, places, products and jargon you actually say. So Caddie's dictation picks them up from what's already there. It knows the people on your invites and the projects in your notes on the first day you use it.

Then it keeps learning. New notes and meetings give it new words, and when you correct something it remembers the fix.

Here is what that difference looks like on one sentence. Say your notes mention a colleague called Meera Iyer, a project called Zephyr and a partner called Kestrel Labs.

The generic engine isn't bad at listening. It hears the sounds fine and picks the closest words it knows. Meera becomes Mira, "loop in" becomes Lupin, Kestrel becomes Castrol. It has no way to know better.

Does it show up in the numbers?

We recorded eight everyday clips: Indian English, Hinglish, tech terms, names, a self-correction, an email address and a number. About 70 seconds of audio in all. We played the same clips into each app through a virtual microphone, so every app heard exactly the same sound. Then we counted the wrong, missing and extra words against what the speaker meant to type.

Caddie's three modes were scored after cleanup, because that's the text you actually get. The chart at the top has the result.

Caddie has three modes. Scribe is the most accurate. Default answers in 0.46 s and makes a few more mistakes. Turbo is for people who would rather fix a word than wait. Apple's built-in dictation is the quickest to start typing and makes about one mistake in every four words.

Speed is the wait plus the fixing

Apps usually compete on how fast the text appears. That's the wrong clock. What you care about is how long until the text is right.

So take a typical dictation of 42 words. Add the wait to the time it takes to fix each wrong word, about two seconds apiece. The order changes.

Apple waits 0.23 s and then leaves you about ten words to fix. Scribe waits a full second and leaves about three. Scribe gets you to correct text in roughly a third of Apple's time.

Default sits fourth here. It waits half as long as Scribe and leaves about two more words to fix. If you mostly dictate short messages, that trade can be worth it. If you dictate notes and emails, use Scribe.

How we made it fast

Accuracy came first, but nobody keeps using dictation that makes them wait. In Default mode, about half a second passes between letting go of the key and seeing your text. Here is where it goes.

1. We tested ten setups on real speech.

   ElevenLabs Scribe both streamed and sent whole, Groq's Whisper v3 and Whisper turbo, Deepgram Nova-3 and Flux, Apple's engine and Parakeet, with Wispr Flow and Kivi as reference. Public benchmarks didn't predict Hinglish at all. One engine quietly dropped the Hindi. Another translated it into English.

2. We stopped streaming.

   Streaming feels fast, but the model has to guess each word before it hears the rest of the sentence. Sending the whole clip when you let go cut Scribe's errors from 13.3% to about 7%. You still see your words live in the notch while you talk, because your Mac transcribes a rough copy on its own.

3. We compress while you speak.

   Caddie encodes the audio as AAC while you hold fn, so the upload is 7 times smaller. For the test set that was 2.3 MB down to 0.33 MB. Scribe's wait went from 1.24 s to 0.91 s, with no loss in accuracy after cleanup.

4. We moved the server to where you are.

   Cloudflare's free plan sent traffic from India through Paris or Marseille, 130 to 260 ms away. A relay in Mumbai, next to Groq's servers in Chennai, cut the round trip from 0.54 s to 0.29 s.

5. One request, and cleanup only when it's needed.

   Transcription and cleanup go out as a single request. Cleanup runs only when there's something to clean: filler words, a change of mind like "no wait, 6 PM", a list, or Hinglish. With cleanup the whole thing takes about 0.45 to 0.6 s. Without it, under 0.3 s.

6. The connection warms up when you press fn.

   By the time you let go, the line to the server is already open.

7. Hinglish stays Hinglish.

   Giving Whisper a list of English names made it translate the Hindi into English, so Whisper gets no list. Your words are matched in afterwards instead, by sound. Scribe takes your words directly, as terms to listen for. If an engine writes Hindi in Devanagari, your Mac turns it into Roman letters, and cleanup never translates.

8. It never fails silently.

   If the cloud is slow, your Mac transcribes on its own. If the cleanup model is overloaded, a backup model takes over.

9. It got cheaper too.

   Default costs $0.11 per hour of speech, down from $0.39.

What this test doesn't show yet

• We haven't yet run the test that isolates the idea in this post: Scribe with and without your words, on the same clips. Until we do, we can't say how much of Scribe's lead comes from knowing your words and how much from the engine itself.

• It's eight clips from one speaker. We're running 100 clips from public datasets next, AI4Bharat Svarah for Indian English and Microsoft MUCS for Hinglish, and we'll publish those numbers when they're in.

• Wispr Flow is a close second on accuracy, less than a point behind Scribe, and it waits less.

• Default trades a little accuracy for twice the speed of Scribe. It isn't the most accurate mode.

• With 70 seconds of audio, a gap of a point or two is within noise. Treat close rankings as ties.

The same goes for meeting notes

A note-taker that lives in your knowledge base knows the people and projects that come up in your calls, what you decided last time and what you still owe. That should make its meeting notes more accurate too, the same way it makes dictation more accurate. More on that in a separate post.