Latency
- Mac
- about 1s
- Windows
- not measured yet
1 process
Here it is next to a heartbeat and a blink.
Not typed out letter by letter.
Each dictation reuses one open connection: no new handshake to wait for.
Stopping the mic, encoding and pasting take about 57 ms. The rest is the network and the model.
Swift on the Mac, Rust on Windows. No browser engine inside.
Transcription runs on MAI-Transcribe 2 in the cloud.
When you aren’t dictating, nothing runs but a once-a-day update check. No audio.
No card, no email, no signup. macOS 14 or later on Apple silicon or Intel, or Windows 11 on an x64 PC. See how blabe and Wispr Flow differ on size, price and privacy.
| Clip | AI cleanup off | AI cleanup on |
|---|---|---|
| 3 s | 450 / 495 ms | 625 / 1,225 ms |
| 10 s | 510 / 915 ms | 705 / 920 ms |
| 30 s | 780 / 855 ms | 1,005 / 1,660 ms |
say, sent through an API harness that mirrors the app’s request path. The everyday figures are 43 dictations made in the Mac app during normal use. Each dictation reused an open connection; a new one adds about 150 ms once.