Speech models on a Mac: measured accuracy and speed
MeetingEcho measured four Whisper models on an Apple silicon Mac: word errors in German, English and French, how long dictated text takes after you stop, and how long an hour of meeting takes. Apple's built-in model is the quick alternative; all of them run on the Mac.
| Model | Size | Word errors (German, English, French) | Dictated text after you stop | One hour of meeting |
|---|---|---|---|---|
| Base | 142 MB | DE13%EN0%FR52% | under 1 s | about 3 min |
| Small | 466 MB | DE2%EN0%FR24% | under 1 s | about 4 min |
| Large v3 Turbo, compresseddictation default | 547 MB | DE0%EN0%FR7% | about 1.5 s | about 6 min |
| Large v3, compressedmeeting default, included | 1.1 GB | DE0%EN0%FR0% | about 2 s | about 13 min |
Measured on an Apple silicon Mac with synthetic voices, which are harder to understand than real people.
Measured with ten synthetic sentences in German, English and French and a 35-second German meeting. The hour is worked out from the 35 seconds and errs on the slow side. All four models detected the language of every sentence. Two full-precision models (1.6 GB and 3.1 GB) are offered but were not measured.
Which model to choose
The included Large v3 model is the meeting default: no word errors in the test, about 13 minutes for an hour of meeting. Large v3 Turbo is the dictation default: text about 1.5 seconds after you stop. Small and Base are faster and need less memory, at the cost of accuracy in French.
Apple's model
Apple's dictation and speech models are built into macOS, need no download, show words as you speak and are the quickest. They cover ten languages with every feature of the app.
Questions and answers
Where do the models come from?
A speech model you choose is downloaded once from Hugging Face; that is the only time MeetingEcho connects to the internet by itself.
How much memory do they need?
Between 0.4 and 2 GB, depending on the model; Settings shows it for each one.
Can I delete a model?
Yes. Settings downloads or deletes each model in one click.
Getting started
A short welcome tour shows you around on the first launch, and Settings start simple. Everything works as it is; Advanced is there when you want more.
Choose your languages.
Up to four you work in, and the one you are learning.
Meet and dictate as usual.
MeetingEcho waits in the menu bar, notices calls, and listens only when you ask.
Read, search and practise.
Transcripts, notes and cards appear on your Mac, ready to search, export or practise.