Back to work

Programmable dictation for macOS

Verbaform

A programmable dictation app for macOS where spoken Verba commands edit the transcript mid-dictation, with an event-sourced Rust core, local Parakeet and whisper.cpp transcription, and a Tauri desktop shell.

Live beta RustTauriONNX Runtimewhisper.cppSilero VADSQLite
Visit
The Verbaform landing hero reading Your voice is the input and the editor, beside an app window with a live transcript, a Listening chip, floating command chips, and a session HUD showing a local Parakeet engine and no network.
Voice as input and editor
The Verbaform command studio showing a spoken command list, a before and after document view, and an execution HUD reporting a local engine and undo state.
Spoken commands, exact results
A command anatomy diagram breaking a spoken command into labeled wake word, function, chain, and scope tokens, above example command chips.
Anatomy of a spoken command
An intent card showing a proposed Verba program, a preview diff with filler words struck out in red, Apply and Cancel buttons, and a note that the document stays hidden from the compiler.
AI writes programs, not documents

Problem

Dictation needs verbatim capture, but voice editing needs deterministic commands that never fire inside ordinary prose.

System

Verbaform keeps the raw transcript as an immutable token layer: every edit is an event, the visible document is a pure fold, and undo and redo are events too. Commands parse deterministically, and AI intent compiles to a visible Verba program that runs only after an approved preview diff.

Highlights

  • Verba, a deterministic spoken command language with 49 functions in a schema-validated registry: trim fillers, redact a name everywhere, chain operations, define macros.
  • An event-sourced core keeps the raw transcript immutable: every operation is an event, the document is a pure fold, and undo and redo are events too.
  • AI stays behind a parser firewall: spoken intent compiles to a visible Verba program, a mandatory preview diff gates execution, and the model never touches the document directly.

Key decisions

  • Keep the grammar deterministic: a spoken command parses exactly or the words stay verbatim text, never a guess.
  • Compile AI intent into a visible Verba program: only the approved program creates events, and the model sees span-only payloads, never the full session.
  • Fail closed on secure input: password fields are detected by independent checks and refused.

Evidence

  • A signed and notarized DMG at tryverbaform.com, with a minisign-verified auto-updater, a fully local free tier, and live Stripe billing.
  • A soak release gate in CI replays 4.3 hours of real prose and asserts zero false commands and zero wake flushes.
  • Property tests hold apply-undo byte-identical, the language core compiles to WebAssembly as a CI gate, and the text-delivery cancel race is model-checked with loom.
  • A public reference for the spoken language at tryverbaform.com/language, and a repo-recorded decision log behind the architecture.