Voice Coding: How to Program by Voice

Programming by voice works in two different ways, and they need different tools. In the first, a command system turns spoken words into precise edits in the code. Talon, with Cursorless in VS Code, is the best-known example. In the second, you dictate prose such as prompts, comments and commit messages to an AI coding tool, which writes the code. The first gives fine control after a learning period. The second needs no new grammar and works in any text box. In both cases, typing speed may matter less than you might expect. In one field study, developers spent about 58 percent of their time on program comprehension, the work of understanding code ¹.

This guide uses voice coding to mean either method. Our overview of voice typing in apps covers the plain dictation side in more detail.

Voice coding means either speaking code commands or dictating to an AI

The Nature feature on voice programming explains the structure: coding by voice needs two kinds of software, a speech-recognition engine and a platform for voice coding ². The engine turns sound into words. The platform decides what those words do to the code. Dictation to an AI skips the platform. The words go straight into a text box, and the AI does the translation into code.

The two approaches differ in what you say, what you learn and where they work.

The rest of this guide takes each approach in turn, then covers the evidence on typing speed and a practical way to combine them.

Command grammars like Talon give precise control, after a steep start

Talon's home page lists downloads for macOS, Windows and Linux (X11) ³, but a 2026 report quotes its developer saying Linux support will soon be removed ⁴. The platform list has changed over time: in 2018, Nature described Talon and VoiceCode as popular platforms that were for Mac OS only ⁵. Check the download page for your system before you plan around it.

The cost is learning. Tavis Rudd built a spoken Python system that required users to learn a complex set of over 1000 commands ⁶. The interviewees in the 2022 study were frustrated by long hours of memorizing unnatural commands to enter code by voice ⁷. The payoff is precision, because each exact command does one predictable thing. The price is a vocabulary you have to build.

How long that takes varies. One software engineer with severe hand pain reported success in two weeks, using four tools ⁸. That is one person's account, not a measurement, but it points to the same advice as the interviewees: start with a few tools and add more later.

If you want to test Talon before you commit to it, a short trial shows you more than any review:

1. Install Talon from the download page for your system.
2. Select a speech engine from the Talon tray icon menu, as the Cursorless setup steps describe.
3. Run a handful of built-in commands and dictation, and note the ones that do what you meant.
4. Pick three edits you make every day, such as selecting a word or deleting a line, and learn only those commands first.
5. After a week of real work, count how many of those edits you now make by voice without stopping to think.

For a shorter answer on how the system works, see how Talon voice coding works.

Cursorless adds structural editing by voice inside VS Code

Cursorless describes itself as a spoken language for structural code editing that lets developers code by voice ⁹. Its overview says every command consists of an action performed on a target ¹⁰. A short spoken phrase names the action and the part of the code it applies to.

Cursorless depends on Talon. Its installation guide starts with Talon and a speech engine ¹¹, and the steps run in this order:

1. Install Talon, then click the Talon tray icon and select a speech engine from the Speech Recognition menu.
2. Install the community Talon commands.
3. Install VS Code and the VS Code talon extension pack.
4. Install the Cursorless VS Code extension, follow the steps for the Talon side of Cursorless, and restart Talon.

If you already live in VS Code, this combination adds the most structure. If you do not, the setup is hard to justify, because Cursorless itself installs as a VS Code extension. The steps above come from that guide as it stood on the access date, so check it again before you install.

Natural-speech tools promise an easier start, but check what they accept

Serenade is the other tool in this group. Its home page says that with Serenade you can write code using natural speech ¹². A 2022 paper describes the product differently: Serenade requires that spoken commands precisely match the commands the system understands ¹³. The paper also says the product was created out of need when a co-founder was diagnosed with RSI ¹⁴. These two descriptions do not agree, and only a test with your own speech will show which one matches your experience.

Does Serenade still work? We could not confirm that. The download is still offered: Serenade's documentation says the app is freely available for macOS, Windows and Linux ¹⁵. But the most recent commit on the main branch of its public GitHub repository is dated July 23, 2022 ¹⁶. We found no shutdown notice, and we could not confirm the company's current status. Before you rely on it, install it on a machine you can test on, and check that it still recognizes your speech. The short answer on Serenade covers the same ground.

Five checks settle most choices before you install anything

Whatever tool you consider, answer these five questions first. Each one takes a few minutes with the vendor's own pages.

• Does it run on your system? Compare the platforms on the download page with the computer you actually use, including the operating system version.
• Is it still being updated? Look for a recent release note or a recent commit on the project's main branch. A date from years ago is a reason to test first.
• Does it work in the app you use most? Cursorless is set up for VS Code. Dictation into an AI tool works in its own chat box.
• How much vocabulary must you learn? A command system needs a set of exact commands. A dictation tool needs you to read the text before you send it.
• Does it send audio off your computer? Some dictation sends recordings to a company's servers for transcription. Read the privacy note for the exact tool.

Dictating prose to an AI coding agent needs no new grammar

This is the approach most developers can start today. You speak a prompt, a code comment or a commit message, and an AI tool does the writing and the code.

Claude Code has dictation built in. Running /voice in the CLI turns it on ¹⁷. Dictation has a privacy cost: according to Claude Code's documentation, the audio is not processed on your computer ¹⁸. For the setup, see our guide to voice input for Claude Code and other AI coding tools, and the Claude Code dictation page.

VS Code offers a similar option. Its built-in dictation converts speech to text in chat, the Agents window, editors and terminals ¹⁹. The page labels the feature experimental ²⁰, so expect changes.

Does spoken text work as well as typed text? A 2026 preprint compared typed and spoken prompts from students. For two of the three problems, students who typed were more likely to succeed on the first attempt than students who submitted unedited voice prompts ²¹. When students edited their transcribed prompts before submitting, the difference disappeared ²². The practical rule follows: dictate the long prompt, then read it before you send it. Students are not professional developers, so treat the result as a guide to habits rather than a measure of professional output.

Dictation also needs care with names. A transcript can spell a function or file name the way it sounds, and the AI will use the spelling it receives. Read every identifier before you send, and fix it in the text, not after the code is written.

Yap, which we make, is an on-device dictation app for Mac and Windows. If you want one tool for every app instead of the built-in button in each one, see the download page. For how speech feeds AI prompts more generally, see our guide to speech to text for AI.

Hand pain and disability shaped the first voice coding tools

The first serious voice coding tools came from programmers who needed them. Talon's developer started building it after severe hand pain ²³. Serenade grew out of a co-founder's RSI diagnosis ¹⁴. The same 2022 paper also describes Tavis Rudd's spoken Python system, which required over 1000 commands ⁶. How many programmers have that need is hard to measure. In the Stack Overflow 2022 survey, 0.35 percent of respondents said they were unable to type or found it difficult to type ²⁴. That share covers people who chose to answer one question on a voluntary survey, so it describes respondents, not every developer's hands. Our guides on dictation for RSI and carpal tunnel and RSI from typing cover the medical side of repetitive strain injury.

Voice has costs too. Before formal research, users of speech recognition software had reported voice problems, from a dry throat to being unable to speak for more than half an hour a day ²⁵. Build breaks into long sessions, and keep the keyboard for work that does not need your voice.

Typing speed matters less in coding than in prose

For coding, typing speed is rarely the bottleneck. In a field study of professional developers, the developers spent about 58 percent of their time on program comprehension, which is the work of reading and understanding code ¹. Faster typing does not shorten the reading, which takes up most of that time.

Touch typing does not explain the difference either. In a 2016 transcription study, touch typists entered text at 57.8 words per minute ²⁶. Typists who had not learned touch typing reached 58.93 words per minute in the same task ²⁷, about 1.1 words per minute more than the touch typists ²⁸.

Speech is not a shortcut for experienced programmers either. In a 2006 evaluation of a speech-based Java programming system, most expert Java developers found it slower than typing ²⁹.

Prose is different. In a controlled study on an iPhone, English speech reached 153 words per minute, against 52 words per minute on the built-in keyboard ^{30 31}. The chart compares the two kinds of entry.

Touch typing shows no advantage in the 2016 test, while speech is much faster for prose. Writing also takes a smaller share of a programmer's week than most other professions. In a 2018 NC State survey, programmers reported spending 26 percent of their work week writing, the lowest share of any field ³². Those are figures people reported about their own time, so read them as rough.

To measure your own speed, time yourself on two samples: a paragraph of ordinary prose, and a block of code you would really write. Compare the words you produce with the time taken, and note how much of the code session went to reading and thinking rather than typing.

The answer to the question is therefore split. Typing speed matters most for prose: prompts, comments, documentation and messages. It matters least for the reading and reasoning that take up most of a coding day. For a shorter answer, see does typing speed matter in coding. To train your speed on prose, see how to type faster and our entry on words per minute.

A hybrid setup of voice, keyboard and AI covers the most work

The motor-impaired programmers interviewed in the 2022 study did not give up the keyboard. All of them used a combination of mouse, keyboard and voice to program ³³. A hybrid is the practical answer, and the tools above fit into it at different points.

A practical order to start:

1. Dictate long prompts, comments and commit messages into your AI tool. Use its built-in dictation, or a system-wide dictation app that works in every field.
2. Read each dictated text before you send it. Unedited voice prompts did worse in the 2026 preprint ²¹, and editing closed the gap ²².
3. Keep the keyboard for exact symbols, short identifiers and quick fixes, and use voice for the longer text.
4. If you use VS Code and want command-based edits, add Talon and Cursorless after the dictation habit is in place, and learn them a few commands at a time. Tavis Rudd's spoken Python system required over 1000 commands ⁶, so start small.
5. Check each tool's status before you build a daily routine on it. For Serenade, that check comes first.

Voice coding is a set of trade-offs rather than a replacement for the keyboard. Command grammars give precision after a learning period. Dictation gives speed for prose, provided you check the words before you send them. For the wider picture of dictation inside apps, see our overview of voice typing in apps, and for the Linux side, see speech to text on Linux.