Voice and tone in a dictation app
How Skald decides what register to write in — and why we won't drift.
A dictation app has to make a choice for the user: do you preserve the literal transcript with all its fillers, or do you clean it up? We chose to clean — but conservatively.
The cleanup model knows the app you’re typing into. In Slack it stays loose. In an email it writes full sentences. In code it leaves your words alone and only converts spoken tokens like “comma” or “newline” to characters.
The rule we keep coming back to: if you didn’t say a name, a fact, or a number, the model won’t add one. When the transcript is ambiguous, we prefer the literal interpretation. The user is the author. The tool is a scribe.