0.1.8
LATEST- Swedish cleanup is now as fast as English. Babbla’s fast cleanup model, S1-mini, has been taught Swedish: fillers go, numbers, dates, times and amounts are written out the Swedish way (“halv tre” becomes 14:30, “tjugotre tusen kronor” becomes 23 000 kr), lists are laid out, and English terms inside a Swedish sentence stay exactly as you said them. Around a quarter of a second per dictation instead of close to a second. The new weights download on first launch (630 MB).
- The Apple Intelligence cleanup option is gone: S1-mini cleans every dictation, and Qwen3 4B is used for translation and stands in until the fast weights have arrived. There is nothing to pick any more.
- Your dictionary is applied before any model sees the text, as exact, sound-based corrections. Names like Quinn, Lynn and Luca are no longer rewritten into dictionary terms that sound alike, and a learned correction can never overwrite an ordinary word.
- Teach a word: name a term in the Dictionary pane, hold the dictation key and say it. Babbla stores how it hears you say the term and corrects it from then on. It also works out on its own how the recognizer mishears your dictionary terms.
- Field glossaries are live from the moment you pick a field. Spoken forms such as “kube control” and “engine x” become kubectl and nginx on every dictation, with no learning toggle and no model needed. Background vocabulary mining no longer waits for Qwen3 4B to be installed.
- A new guided setup: thirteen short pages that take you from permissions to your first message, email and list, with real download progress measured from disk and a note while models compile for your Mac. Setup waits until both permissions are granted and tells you what to flip if the microphone was turned off earlier.
- History now compares what was pasted against what was heard, so a dictionary correction shows up as the change it is.
- Insights shows real app icons, and the hidden pill no longer keeps a CPU core busy: Babbla idles at 0% again.