The Overlay

The overlay is always on screen. At rest it is a short pill — Ready, a five-bar mark and the tally — sitting centred about 12 pixels above the bottom of your work area, clear of the Dock, the Windows taskbar or a Linux panel.

It never changes layout when it could change colour instead. The tally dot, the target-app icon and the meter hold their positions through every transition, so nothing jumps in your peripheral vision while you are looking at another window.

Colour has exactly one job: how loud you are. It is a stepped ramp, not a blend, so the threshold stays readable.

• Signal — quiet. Ordinary conversational speech sits here, near −20 dBFS.
• Warm — loud. Starts at about −17 dBFS.
• Hot — peak. Starts at −6 dBFS.

Which is why you only ever need to learn one thing: orange means the microphone is hot — legible across a room, on a machine you are not looking at.

The meter is not a pulse. A new bar is drawn every 80 ms and they scroll left as you speak, oldest on the left and newest on the right, beside a running clock. Twenty bars fit beside it, eighteen once the dictation passes ten minutes and the clock widens to minutes and seconds. A pause between words is a dip you can see rather than something you infer.

Bars rise in 20 ms and fall over 250 ms. A meter that follows amplitude sample-for-sample looks like noise; one that averages looks dead. The fast attack catches the transient of a consonant and the slow release lets the bar fall through the gaps between syllables. The bars are drawn against a −50 dBFS floor rather than the −60 dBFS the level arrives on, so room tone sits flat on the baseline instead of shimmering.

Tip: if the bars stay flat while you talk, the wrong input device is selected. Check Settings → General → Input device.

Dictation lands in whatever last had focus. The icon between the tally and the meter is the only thing that confirms which window that is before you commit thirty seconds of speech to the wrong one. When an application has no icon the slot keeps its size and shows the first letter of its name, so the pill never reflows as focus moves.

Each of these is the live overlay component, given a different state.

• Resting

  Push-to-talk. The microphone is closed. The mark, in Dim, and nothing else.

• Listening

  Hands-free. The same mark in Signal. Not a second layout — one question answered.

• 

  Recording

  Hot tally. A scrolling waveform and a running clock, a new bar every 80 ms — a pause between words is a dip you can see.

• 

  Silent

  The same pill, the bars resting on the baseline. The microphone is open and hearing nothing — a muted input, or the wrong device selected.

• 

  Transcribing

  Warm tally, eight stepped spokes, and no words. It used to carry a stage label, a language line and a shimmer bar: three things narrating one fact.

• Composing…

  Composing

  Signal tally, a drawn zap and the word Composing. You are holding the Compose key: say what to write, or how to change the text you selected.

• Couldn’t insert into Slack

  Retry

  Couldn’t insert

  The only card. It names the window that refused the text; Retry tries again and Dismiss puts it away.

It reports; it does not repair. There is no Backtrack, Copy or Correct button — the insertion-failure card's Retry and Dismiss are the only buttons it carries. And there is no “inserted” state: when the text lands, the pill goes straight back to rest rather than holding a word count for something you have just watched arrive in your own document.

Related: Push-to-Talk Setup, Compose and Troubleshooting.