SaypadHow accurate is offline dictation on a Mac?
Offline dictation means the speech model runs on your Mac instead of a server. The catch is that the good models are big. We wanted to know how much accuracy you give up with a smaller one, so on 8 October 2026 we ran 32 real dictations through four versions of OpenAI's Whisper and compared the text word by word.
The short answer
The 874 MB model gave us the same text as the full 1.6 GB one in 31 of 32 dictations, and the one difference was punctuation. The 190 MB model is about two and a half times faster, but roughly every fifth word came out different. Going smaller still doesn't work: in an earlier test the tiny and base models turned Russian into nonsense.
What we tested
32 real dictations taken from the app's history, each 5 to 18 seconds long, a bit under 6 minutes of speech in total. Most were Russian with English words mixed in (product names, code terms), which is harder than clean English. Everything ran on a MacBook Pro with an M4 Pro chip, using the same engine Saypad uses (whisper.cpp with Apple's Metal graphics).
There's no human-typed transcript for these recordings, so we used the biggest model as the reference and counted how many words each smaller model got differently. That tells you how much you lose by going smaller. It doesn't tell you how often the big model itself is wrong.
Results
| Model | Download | Words different from the best model | Same text, word for word | Time per dictation |
|---|---|---|---|---|
| large-v3-turbo (full, f16) | 1,625 MB | reference | – | 2.14 s |
| large-v3-turbo q8_0 | 874 MB | 0.0% | 31 of 32 | 1.98 s |
| large-v3-turbo q5_0 | 574 MB | 2.4% | 21 of 32 | 2.21 s |
| small q5_1 | 190 MB | 22.0% | 0 of 32 | 0.84 s |
Time per dictation is a cold start: the model is loaded from disk for every recording. In the app the model stays loaded while you use it, so a 7.6-second sentence appears about 1.2 seconds after you let go of the key.
What this means in practice
Short messages and notes are fine with the small model. It gets the gist and most words right, and it's very fast. You'll fix a word here and there.
Longer text, names and technical terms need the large model. That's where the small one slips most: it hears a product name as an ordinary word or drops an ending.
The 8-bit version is the one to have. It's half the size of the full model and we couldn't find a difference in the text. The 5-bit version saves another 300 MB, but it changed words in 11 of 32 dictations and wasn't any faster.
How Saypad uses this
Saypad's installer includes the small model, so you can dictate a minute after installing, even offline. The 8-bit large-v3-turbo then downloads in the background over Wi-Fi, and Saypad switches to it by itself once it's ready. You can stop that download in Settings if disk space matters more to you.
Two things help accuracy with any model: speak in whole sentences rather than single words, and add names and jargon to Settings → Vocabulary.
Limits of this test
One speaker, one microphone, one Mac, and mostly one language. We make Saypad, so we had a reason to pick the model that works best for our app. That's why the method is spelled out above. Run your own recordings if your use is different.
Questions
Is offline dictation as accurate as cloud dictation?
We didn't compare against cloud services in this test, so we won't put a number on it. What we can say is that large-v3-turbo, the model at the top of the table, runs entirely on an Apple Silicon Mac and needs no connection once downloaded.
Which Whisper model should I use on a Mac?
If you have the disk space, large-v3-turbo in its 8-bit version (874 MB). In our test it produced the same text as the full 1.6 GB model and was slightly faster. The small model is fine for quick notes, but it gets roughly one word in five different.
Do I need an internet connection?
Only to download a model. After that, dictation works with Wi-Fi off. Saypad ships with the small model inside the installer, so even the first dictation works offline.
Does it work with languages other than English?
Whisper detects the language on its own. Most of our test recordings were Russian with English technical terms mixed in. In an earlier test the smallest models (tiny and base) turned Russian into nonsense, while small and larger held up.
Can I check these numbers myself?
Yes. The method is described below, and the technical write-up on FnFlow, our edition for developers, lists the exact commands. Your results will differ with your voice, language, microphone and Mac.