MyWare Voice — dictation

Press. Speak. It's typed.

Voice is push-to-talk dictation that lives on your machine. Press your keybind, say what you mean, press again — the transcript goes to the clipboard and pastes into the window you were working in. No account, no cloud service; the audio never leaves the device.

How it works

Three steps, on repeat.

  1. Press your keybind

    A small listening indicator appears with your live level. The model is already warm in memory — capture starts immediately, load time was paid on the first utterance of the session.

  2. Speak

    NVIDIA Parakeet TDT 0.6B v3 (int8) recognises 25 languages with punctuation and capitalisation built in — no separate punctuation model, no post-processing service.

  3. Press again

    The text is tidied by a rules pass — filler cleanup, spoken punctuation, "scratch that" to discard, snippets — then copied to the clipboard and pasted into the window the capture started in. If focus moved, it copies and tells you instead of pasting blind.

Between speech and paste

Rules, not magic.

Every transcript passes through a table of text rules you can read and edit. A rule either fires or it doesn't — which is why it cannot invent a word you didn't say.

  • CleanupRemoves fillers ("um", "uh"), collapses repeated words, tidies spacing and punctuation.
  • Spoken punctuationOptional "comma", "question mark", "new paragraph" → the real characters, pasted as text.
  • Formatting"at sign" → @, "dot com" → .com, so addresses come out written, not spelled.
  • Commands"Scratch that" discards the take; "as a list" reshapes it into bullets; snippets expand a trigger phrase into fixed text.
  • StreamingA second keybind pastes phrases as you speak, committing at pauses — an optional word-by-word engine (Pro) narrows the delay further.

Optional & local-only

An assistant that never leaves the loopback interface.

If you run Ollama on this machine, Voice can read your (opt-in) dictation history and suggest the correction rules you keep making by hand. The client accepts only 127.0.0.1, localhost or::1 — a LAN or cloud host is an error, not a configuration. Nothing is applied without a click, and the model is never in the path of a paste.

Measured, not promised

Fast enough to keep up with talking.

On a Ryzen 7 PRO 8840U (4 threads, int8), a 45-second clip decodes in about 2.2 seconds — roughly 20× realtime — with the warm model holding about 1.5 GB of memory. The daemon times every delivery from keypress to paste and reports it, so the number that matters is one you can see, not one we chose.

Platforms

Linux first, honestly labelled.

Voice works end to end on Wayland (sway is the reference compositor), including streaming dictation. The Windows build compiles and is type-checked but has not been run on real hardware yet — when it ships here, it will say so on the tin. It needswl-copy and a Wayland session on Linux; X11 is not supported.

Free & Pro

Everything that dictates is free.

The daemon, the models, the settings window, streaming by phrase, history, the tray — free. Pro is a signed licence file, checked offline against a public key in the binary; it unlocks the extras, starting with the word-by-word streaming engine. No fetch, no phone-home: an expired licence drops back to free with a note, never a lockout.