Voice to text for coding

What voice is actually good for in coding

Start with the honest split. Code syntax, the brackets, the indentation, the exact variable names, the punctuation that has to be character-perfect, belongs to the keyboard. A microphone will always be a worse tool for typing a semicolon than the semicolon key. No app changes that, including ours.

Now the other column, the one that's secretly most of a developer's day. Comments. Commit messages. Pull request descriptions. Design docs. The Jira ticket you've been avoiding. The Slack reply explaining why the build is red. The reply to a code review. None of that is syntax. All of it is natural language, the exact thing speech recognition is built for. You speak it at conversation speed and read it back instead of typing it at keyboard speed.

So the rule of thumb is simple: if what you're writing would read fine in an email, dictate it. If it has to compile, type it. Most coding tools sell you the dream of writing functions out loud. I'd rather sell you the boring win: the prose around the code, which is more of your day than the code is.

The 150-versus-40 gap

Here's the number that makes the case. Most people speak around 150 words per minute in normal conversation. Most people type somewhere around 40 words per minute, and that's on a good day with no thinking pauses. That's not a small edge. For any block of plain-English text, talking is close to three and a half times faster than typing it.

On a comment or a commit message, that gap is a few saved seconds, which doesn't sound like much. It adds up over a day, but the real payoff isn't the seconds. It's that you write the documentation you'd otherwise skip. The comment that explains the weird workaround actually gets written, because saying it costs ten seconds instead of a minute of typing you'll talk yourself out of. The friction was never the words. It was the keyboard between you and the words.

Dictating prompts to your AI assistant

This is the one I'd actually argue for. The way most of us write code in 2026 involves typing paragraphs at an AI assistant: Claude Code in the terminal, a chat panel in your editor, GitHub Copilot Chat, Cursor. "Refactor this function to take an options object instead of four positional arguments, keep the existing return type, and add a JSDoc comment explaining the cache behaviour." That's a sentence you'd speak in eight seconds and type in forty.

And prompts are forgiving in a way code isn't. The assistant doesn't care if you said "um" or restarted the sentence; it reads intent, not punctuation. So the one place voice is unambiguously a win in a coding workflow is the prompt box. You're already writing English there. You're already not worried about exact syntax. You speak the instruction, it lands as text, you hit enter. If you take one thing from this article, it's this: voice for prompts, keyboard for syntax. The wpm math backs it — about 150 spoken against 40 typed.

The fastest way is a system-wide hotkey

Here's the mechanism that makes this work in any editor and any terminal without a plugin per tool. Whisper by Remskill is not a VS Code extension, a terminal plugin, or a browser add-on. It's a desktop app that behaves like a keyboard: press a hotkey, speak, and the transcript is pasted at your cursor: in your editor, your terminal, your AI chat panel, your issue tracker, anywhere text goes. Because the OS thinks you're typing, it works the same everywhere.

Setup is short:

That's the whole loop. The text appears where your cursor was, you read it, you keep working. No window to switch to, no clip to scrub through.

Speak, and the text lands in your editor

Once it's running, the experience is unremarkable in the best way. You put your cursor in a comment line or a commit field, hold the key, talk, let go. A second or so later the text is sitting there as if you'd typed it. The flow goes from "stop, type, fix the typos" to "speak, done."

Because local transcription runs on your machine (pure-Rust, no Python sidecar, no server in the loop) it works offline. For a developer, that's not a footnote. Your proprietary code, the internal API names in your comments, the bug description that mentions a customer: none of it has to leave your laptop for a dictation feature to work. We'll come back to why that matters for proprietary work.

Cleaning up filler with AI

Spoken language is messy. You say "um," you restart sentences, you trail off. Whisper has an optional AI enhancement step that trims the filler and tidies the phrasing before it pastes. So "uh, yeah, so this, this function, it like, handles the retry, the retry logic" becomes "This function handles the retry logic." For a commit message or a doc comment, that's the difference between dictated noise and something you'd actually commit.

That cleanup runs locally through Ollama, free, on your own machine. Pro users can route it through the cloud instead, but the filler-cleanup benefit doesn't require Pro. It's in the free local pipeline. You can also turn it off and paste the raw transcript, which for a quick AI prompt is often exactly the right call, since the assistant doesn't mind the "um" anyway.

What the full Whisper app looks like

The hotkey is the part you'll use most, but there's a settings surface behind it. You pick your transcription engine: Whisper models, whose multilingual variants cover 99 languages and whose English-only .en builds cover exactly one, or NVIDIA's Parakeet, about 600 MB, 5 to 10 times faster than Whisper on CPU, covering English plus 24 European languages. You set your hotkey, manage history, save presets, and add custom vocabulary, which is handy when your codebase is full of words no dictionary has heard of. None of that is required to dictate one comment. It's there when you want to tune it.

When to reach for Talon or Serenade instead

Now the honest part, the section AI-written articles never include. If what you actually want is to write raw code by voice, to navigate the file, edit syntax, refactor, and never touch the keyboard, then Whisper is the wrong tool, and so is any general-purpose dictation app. You want a tool built for hands-free coding, and two stand out.

Talon Voice is the heavyweight for full hands-free control. It does voice commands, noise control (clicking with pop and hiss sounds), eye tracking for the cursor, and deep customization through Python scripts. It runs on Windows, macOS, and Linux. People with serious RSI run their entire machine through it. It has a learning curve, and a Patreon tier for early features and priority support, but for true keyboard-free operation there's nothing else like it.

Serenade is the gentler entry point. It's an open-source, free voice assistant built specifically for writing code, with a speech-to-code engine made for developers rather than general dictation. It plugs into VS Code, IntelliJ, PyCharm and others, can run locally or in the cloud, and handles the common languages: Python, JavaScript, TypeScript, Java, C and C++, and more. If your goal is "edit syntax by voice without a steep curve," start there. If your goal is "dictate the comment, the commit, and the AI prompt while my hands stay on the keyboard for the code," that's Whisper. Different jobs. I'd rather point you at the right one than pretend we do everything.

If you want hotkey dictation rather than full voice coding and you are on a Mac, VoiceInk is the other name worth weighing; the honest VoiceInk alternative comparison lays out where each one fits a developer day.

What Whisper costs

The local dictation pipeline (transcription and the AI cleanup over Ollama) is free for any signed-in user, with no card at signup. So dictating your comments, commits, and AI prompts with Whisper costs nothing. Whisper Pro adds the cloud features (OpenAI transcription, cloud AI enhancement, voice web search) on a separate trial. The exact numbers live on the pricing page rather than here, because prices move and a blog post is a bad place to keep them current.

Voice will not replace your keyboard for code, and anyone who tells you it will is selling something. What it replaces is the typing tax on everything that isn't code: the comments you skip, the commit messages you keep short to save effort, the prompt you'd rather speak than peck out. Aim it there and it earns its keep. Aim it at a for-loop and you'll throw the headset across the room, the way my relative did with Dragon on a Windows 98 machine. Some things take twenty-five years to get right. Dictating the prose around code is one we finally have.