Gwen

The challenge

Dictation tools for the Mac are either tied to one app or send your voice to a server. I wanted to hold a key in any app, speak, and get clean text at the cursor, and to translate a selection or what I say, without audio or text leaving the machine.

My role

Sole developer, end to end. I designed the product and built the macOS app, the listener, the punctuation rules, the translation flow, and the marketing site.

Constraints

  • Local by default. Speech-to-text, punctuation, and translation run on the Mac. The one step that can leave it is opt-in and off.
  • Any app. There is no plugin per editor, so the text has to land in whatever field has focus, including terminals.
  • Never worse than what you said. Cleanup can add marks and fix spellings, but it cannot rewrite the take.

Key decisions

  • Two processes. A Swift app owns the keys, the microphone, the bottom bar, and the paste. A Python listener owns segmenting, Whisper, and text cleanup. The app writes 16 kHz audio into a FIFO and the listener reads it.
  • Punctuation is rules, not a model. Spoken commands such as “new paragraph” come first, then each sentence gets its marks from its wording, in English and Spanish. A text no rule touches comes back byte for byte.
  • Paste where the cursor is, and say so when it can’t. The clipboard is saved, used for the paste, and restored. With no text field in focus, the bar shows the text as copied instead of dropping it. In a terminal, line breaks are flattened so a dictation never runs a command.
  • Learn from fixes. After a paste Gwen reads the field back while you are in it. The same spelling fixed twice becomes a rule, stored in a private file under ~/.gwen. Nothing is logged or sent.
  • Translation uses Apple’s on-device framework, so it needs macOS 15 and no API key. Language detection is limited to Gwen’s six languages, so a short Spanish phrase is not misread as another language.
  • Agent polish is opt-in. A local agent CLI can tidy each take, with no tools and a cost cap per take. If it is missing, slow, or changes what you said, the rules result is kept.

The marketing site

Gwen has a site at gwen-chi.vercel.app. It is one static HTML file that Vercel serves as is, with no build step. The hero plays the bottom bar through a dictation, a translation, the wake word, and a hands-free take, so you see what the app does before you install anything.

The Gwen site’s hero: “Dictate and translate in any app”, above a demo of the bottom bar over a Notes window

The features section covers what is not obvious from one take: the keys, the language chip on the idle bar, the words Gwen learns from your fixes, and “Hey Gwen”. Each card shows the piece of the bar it describes instead of an illustration.

The features section of the Gwen site, with cards for voice input, the language chip, learned words, and the Hey Gwen wake word

The site is also how Gwen is installed. It serves the script behind a one-line install that clones the repository, builds the app, and starts it at login, and it links a disk image for Apple silicon and Intel.

The install section of the Gwen site, with the one-line curl command and a Download for Mac button

Results

Gwen is open source under the MIT license. It dictates into any app with a held key, a double tap, or “Hey Gwen”, translates a selection in place or what you say, and exposes a gwen:// URL scheme so other apps can start a dictation or a translation. It installs with one line from its site, gwen-chi.vercel.app.

Tech stack

Built with Swift and AppKit, Python, whisper.cpp, Apple’s Translation and NaturalLanguage frameworks, and the macOS Accessibility API.