Source: https://fnflow.app/voice-coding-mac

# Voice coding on a Mac: what actually works
"Voice coding" covers two very different setups. One replaces the keyboard: you speak commands, symbols and identifiers, and a grammar turns them into keystrokes. The other leaves code to the keyboard or an agent and uses speech for the writing around it, such as prompts, specs, review comments and commit bodies. Most developers who try voice on a Mac want the second and end up evaluating tools built for the first, or the other way round.

## Two things called voice coding

The first kind is command-driven. You learn a vocabulary of spoken commands, and each one becomes a keystroke, a symbol, a formatted identifier or an editor action. It's the route people take when they need to stop typing altogether, and it takes practice, like learning a new keyboard layout.

The second kind is dictation. You talk in sentences and get sentences back, with punctuation and capitals. It doesn't help you type user_id or a closing bracket, but it covers a surprising share of a developer's day, because much of what we type isn't code.

Both are useful. They just solve different problems, and picking the wrong one is the usual reason people give up on voice after a week.

## Speaking syntax: command tools

Talon is the best-known tool of this kind. Its wiki calls it a hands-free input replacement for the keyboard and mouse. The site shows spoken commands like "select all" and "delete word", custom commands written in files, cursor control by eye tracking, and mouth sounds mapped to actions ("pop to click, hiss to scroll"). It runs on macOS, Windows and Linux on X11, and the wiki recommends installing a community set of command files to start from.

Editors are adding smaller pieces of this. VS Code's terminal dictation, for example, removes ordinary punctuation and converts spoken symbol names into symbols, because a shell command needs different output from an email. That helps for short commands. It's not a replacement for a full command grammar.

If your goal is to write and edit code without touching the keyboard, look at command tools first. A dictation app won't get you there.

## Speaking prose: where dictation fits

Look at what you typed yesterday that wasn't code. For many developers it's mostly prose:

- Prompts for Cursor, Claude Code or Codex, often a paragraph with the symptom, the suspect file and the constraint.
- Pull request descriptions and review comments.
- Commit message bodies that explain why a change was made.
- Specs, design notes, tickets, and replies in Slack.

This is where dictation pays off. Speaking a paragraph is faster than typing it, and people tend to explain more when they talk, which is what an agent or a reviewer needs. Agent-driven work shifts the balance further. If an agent writes the code, your job is describing what you want, and that's prose.

## A workflow that holds up

- Say the intent and context out loud: what should change, where, what must stay as it is and how to check it.
- Type the exact parts: file references with @, flags, error messages and identifiers. Paste stack traces instead of reading them.
- Teach the recognizer your names. Claude Code's own voice mode adds the project and branch name as hints, and most dictation apps have a vocabulary list.
- Read before you press Enter. A misheard identifier sends an agent to the wrong place, and it won't complain.

The split matters more than the tool. Speak the sentences, type the tokens.

## Where FnFlow fits

FnFlow is a dictation tool, so it covers the second kind. You hold fn (or right ⌘), talk and let go, and Whisper transcribes the whole take on your Mac. The text is pasted at the cursor in whatever app you're in, from the Cursor chat to a terminal or a GitHub comment box. It keeps your words as you said them, without an AI rewrite, and every dictation is saved with its audio in case a paste goes astray.

The vocabulary list takes up to 60 terms as spelling hints. In one test sentence, five terms in the list turned "update Zapot" into "update Saypad". Replacements fix words that keep coming out wrong. More on that in [custom vocabulary for Whisper](whisper-custom-vocabulary).

What FnFlow doesn't do is commands. It can't move the cursor, select a word, format snake_case or press keys for you. If that's what you need, a command tool is the right choice, and the two can live on the same Mac. For the prompt side, see [how to dictate a good prompt](dictate-ai-prompts) and [voice dictation for Cursor](dictation-for-cursor).

Working through an agent instead? See [vibe coding with voice input](vibe-coding-voice-input).

## Questions

**Can I write code by voice with a dictation app?**
Not well. A dictation app turns speech into sentences, so "open paren" comes out as words. Writing syntax by voice needs a command tool such as Talon, which maps spoken commands to keystrokes. Dictation apps are good at the prose around code.

**What is Talon?**
Talon describes itself as a hands-free input replacement for the keyboard and mouse. You speak commands like "select all" or "delete word", write your own commands in files, and can use eye tracking and mouth noises. It has downloads for macOS, Windows and Linux (X11).

**Is dictating prompts to an AI agent really faster?**
For prompts with real context, usually. Speaking a paragraph about the bug, the suspect file and the constraint is quicker than typing it, and people tend to include more detail when they talk. Short commands are often faster to type.

**Does FnFlow understand programming terms?**
Whisper knows common ones. For your own names, add them to Settings → Vocabulary as spelling hints, and use replacements for terms it keeps getting wrong. Paths, flags and exact identifiers are still safer typed or pasted.
