How Cleft's transcription got fast, private, and accurate
Cleft has transcribed your voice on-device since the first builds; the cloud only handled the summary step. Why we chose that early, and how we made it fast.
Founder notes
RSSBehind-the-scenes technical notes, founder notes, and build logs.
AuthorsJonathan and Justin
ShapeBuild notes, tradeoffs, and OSS work
CadencePublished when there is a real build story
7 entries
Cleft has transcribed your voice on-device since the first builds; the cloud only handled the summary step. Why we chose that early, and how we made it fast.
CarPlay is the most constrained surface we've shipped: no screen, templated controls, and signal you can lose at speed. How those limits shaped the feature.
1.12 through 1.12.3 added almost no new buttons on purpose. We spent the quarter making the app you already use harder to break.
The hard part of a voice app is clean audio on real devices that connect late, switch mid-sentence, and drop calls. Here's how we made recording survive that.
People wanted a black hole to dump a thought into and trust it synced everywhere. Sync had drifted from that, so for 1.10.1 we rebuilt the sync layer.
Justin published Liquid Speech: an MIT-licensed Flutter package for Apple's native SpeechAnalyzer API that still compiles cleanly on older iOS and macOS.
react-native-skia wouldn't compile for Mac Catalyst during an early Cleft build. I revived a stalled community fix and got it merged upstream in v2.3.0.