Say “kubectl apply dash f deploy dot yaml” out loud and most dictation tools hand you a sentence, not a command. Dictating technical text is where general speech models fall apart, because the vocabulary is not English.
“PyTorch”, “gRPC”, “useEffect”, “--force” — none of these are words a language model expects, and its instinct is to “correct” them into something plausible. The fix is not a bigger model; it is telling the model what world it is in.
Sona lets you teach it your vocabulary — function names, product names, CLI flags, framework jargon. Those terms are fed to the recognizer as context, so “kubectl” lands as one token instead of three guesses, and “dash dash force” becomes “--force” instead of prose.
When you dictate into an editor, Sona can read what is already on screen and use it as a hint — the file you are in, the symbols around your cursor. A name you used three lines up is far more likely than its English near-homophone, and now the model knows that.
It is not magic; it is grounding. Give the recognizer the same context a developer has and it stops guessing.
Dictated code is a fast first draft, not a blind commit. Sona gets you to “compiles and reads right” with your hands off the keyboard — the thinking is still yours.