Dictation for macOS
What you say stays on your Mac.
Tap a hotkey, talk, and your words land at your cursor, in any app. Wavelet runs the latest open source speech models on your own machine, inside a harness built to work fast and accurately everywhere you do.
Your 30-day trial starts with your first successful dictation. For M series Macs. $29 one time, including one year of updates; yours forever. The beta is open while Wavelet is in testing.
The beat before the text lands is 450 ms, within the short-take range measured for the fast models on our test Mac.
Why Wavelet
- A common miss
Tell Anna East the Cooper Netties rollout moves to Thursday.
WaveletTell Anaïs the Kubernetes rollout moves to Thursday.✓
Accurate
Dictation you do not have to go back and fix. Your Mac runs models chosen for accuracy rather than serving cost, and each new generation of open models ships only after it wins on hard real recordings. - $29 once. one year of updates included
No subscription
Pay $29 once, get one year of updates, and keep what you have forever. If a new year of work earns it, you can choose to buy another update year. Nothing renews automatically. - Learned from your corrections Saguenay✓ TensorRT✓ Marchetti✓No list to maintain
Better over timeIn progress
Nobody should hand-build a dictionary of their own vocabulary. We are building a loop where a local model studies your corrections, privately, and teaches the dictation your names and jargon on its own. - Wi-Fi
Private and offline
By default, Wavelet transcribes on your Mac, so a dead zone changes nothing. Your audio and transcripts stay there, with no Wavelet account or sign-up.
Measured on your voice, not on a benchmark
Read a dozen short passages and Wavelet scores every installed model on how it hears you, not how it averages across everyone else. Accents move these numbers more than marketing admits; that is exactly why the measurement is yours.
The quarterly report shows revenue of three point two million, up eleven percent from last year.
The MRI and the EKG both came back clear, but the CBC flagged a slightly elevated white-cell count.
I rebased the feature branch onto main, but the CI pipeline is still failing on the auth middleware.
The passages are from Wavelet's real calibration set; the illustration compresses about three minutes of reading, and your numbers will differ. English first in this release. Several of the models can hear other languages, and we are not supporting that yet.
Next we are exploring dictation for healthcare providers, where accuracy and privacy are not optional. If that is your world, tell us what you need.