Skip to content

Glossary

Voice dictation (for coding agents)

Last updated

Definition

Voice dictation is speaking text instead of typing it, with speech recognition turning your words into text in the field you're typing in; for coding agents such as Claude Code or Codex, it means speaking a prompt into the agent's message box.

Also called: speech-to-text prompting, dictating prompts, voice input for Claude Code.

macOS has dictation built in, and on recent Macs its speech recognition can run on the device itself. General-purpose dictation is tuned for everyday language, though. Prompts for coding agents are full of words it doesn't expect: file names, flags such as --help, identifiers such as userId, your repository's name, and skill commands such as /review.

Dictation for coding agents adds a step between hearing and inserting: fillers and stutters come out, spoken symbols become symbols ("slash review" → /review), numbers become digits, and the project's own names are spelt the way the code spells them. Some tools also have a language model rewrite the transcript into a tidier prompt before it goes in.

Long, specific prompts get better results from coding agents, and speaking them is quicker than typing them. The trade-off is review: what was heard should be easy to check before it's sent.

In Pairbar

After insertion, rate the dictation with thumbs up or thumbs down. Thumbs down opens the words: choose one word or the first and last words of a phrase, then save a replacement with Learn correction. Corrected words are marked green. Choose Add another to fix the remaining words or phrases. Your personal dictionary remembers replacements for future dictations; you can remove it in Voice settings. Ratings and the original inserted text stay in your private history.

Pairbar dictates into Claude's and Codex's message boxes and its own composer: hold ⌃⌥Space to talk. Speech is recognised on your Mac by macOS's own models, then cleaned up locally by mode (Prompt, Message, Verbatim, Code, Notes, Translate to English), with your project's names and your personal dictionary as hints. Optional Polish has your own Claude or Codex rewrite it, on your plan. Nothing is sent to the chat until you send it. See Voice dictation and dictating into Claude Code.

Frequently asked questions

Does voice dictation send audio to the cloud?
It depends on the tool. macOS can recognise speech on the device. Pairbar uses that on-device recognition and keeps no audio unless you turn that on; only its optional Polish sends the text to your own Claude or Codex.
Why does ordinary dictation get code wrong?
It's tuned for everyday speech, so identifiers, flags and project names come out as ordinary words. Tools built for coding prompts add those words as hints and convert spoken symbols.

Related