Lipflow is a local Mac tool that watches your lips and types the words you silently mouth, an open-source alternative to Wispr Flow

Dictation stops working the moment you cannot be heard. Lipflow is a Mac app, released as open source by GitHub user amywork777, that reads the camera instead of the microphone: hold a key, silently mouth the words, and Lipflow types them at the cursor. The project passed 163 stars within three days of its first public commit on 29 September 2026.

Lipflow is MIT-licensed and runs on an Apple Silicon Mac on macOS 13 or later, with the Liquid Glass look on macOS 26. It needs roughly 2 GB of disk. Everything personal, from camera clips to trained models and learned phrases, stays in a single folder under Application Support. The author's name for the category is "Wispr Flow for your lips," a reference to Wispr Flow, the paid cursor-dictation tool many developers already use.

How it works

A lip-reading model watches short clips of the user's face, taken from the Mac's camera, and returns its best guess at what was said. First-pass accuracy is high enough to read everyday news footage: Lipflow's README shows raw model output that reconstructs sentences like "Born in New York City, and raised mostly in Chicago, Nancy Davis graduated from Smith College in 1943" close to verbatim. The raw output is upper-case, flat-cased, and misses punctuation.

Lipflow then runs an optional LLM cleanup pass that picks between the lip reader's top three guesses using the user's recent dictation history, then fixes capitalisation, punctuation and numbers. The pass is optional because the first guess is often correct, so the project author has designed around the common lip-reading failure: words that look the same on the lips, like p, b and m, or f and v, which otherwise turn "while in office" into "wallet officer."

Setup and training

The first launch runs an eight-minute setup that trains the model on the specific face in front of the camera. The project encourages re-training: when a user types a correction within thirty seconds of Lipflow pasting the wrong word, Lipflow saves the clip with the corrected sentence and uses it in the next practice round. The correction loop is on by default and can be switched off in settings.

Lipflow also reads Wispr Flow's local database if the user has it installed, in read-only mode, and saves the user's existing dictation phrases to a local phrases file. Nothing leaves the Mac. The imported phrases are then used to pick between the lip reader's top five guesses using a small per-user model of word pairs, and to prime the cleanup model with past sentences similar to the current one.

Where it fits in a working day

Silent dictation covers the cases spoken dictation does not: a crying baby in the next room, a coffee shop table, an open-plan office, or a late-night session with a sleeping partner. The author concedes in the README that silently mouthed speech is harder than filmed speech because the lip movements are smaller, so the user should expect more errors on a webcam than on news footage. The LLM cleanup pass is where those errors are supposed to get caught.

For a working developer or writer, the practical question is whether Lipflow can learn fast enough from corrections to be useful before the novelty wears off. The project is three days old. The code and the training loop are public, so the next round of improvement is visible.

Source

• GitHub, amywork777, Lipflow, first public commit 29 September 2026.