Voice notes in. Clean text out.

Speaking is easy. Everything after it takes time: re-listening, typing, tidying, deciding what mattered. This is a placeholder case study for a small utility that turns a voice note into something you can use.

The problem

A transcript is not the same as a useful note. The transcript keeps the false starts, the repetition, and the parts where you changed your mind. A useful note keeps the decision and the next action.

The first version of the product gave people a wall of text and called it done. It was accurate and almost useless.

What people actually did

We watched people record, generate, and then edit. Most of them cut the output down to a third of its length before sending it anywhere. The editing was the product, and we had hidden it behind the generation.

The redesign

We kept one primary action, then gave the result a shape. People could accept the summary, expand the detail, or send the whole thing back for another pass.

Testing it

Four people tried the new flow with their own voice notes. The feedback was specific: the summary was too certain, the follow-ups were too eager, and the tone did not sound like them.

So we added a control for tone and a way to mark something as “not a follow-up.” Neither addition was clever. Both made the output feel like it belonged to the person who recorded it.

“

I stopped rewriting the note. I just fixed the two things it got wrong.

”

The result

The placeholder metrics below are here to show the layout. Replace them with real results.

• Median time to a sendable note dropped.
• Fewer people abandoned the result screen.
• More notes were shared instead of discarded.

Takeaways

When a tool does the tedious part, it still has to leave the person in charge of the meaning. The best interface for AI is often a good review screen.