Skip to content

Writing code by voice

Short answer: prose dictation is bad at code, because code is mostly symbols and identifiers rather than sentences. YazSes has a dedicated Spoken Code Mode that converts spoken symbols into punctuation and spoken word-groups into cased identifiers, plus voice-driven git and a safety gate for the terminal. All of it runs offline.

Why ordinary dictation fails at code

Dictate "def get user name open paren user id close paren colon" into a normal speech-to-text engine and you get exactly that — an English sentence. What you wanted was:

def get_user_name(user_id):

The engine has no way to know that "open paren" is a symbol rather than a word, or that "get user name" is one snake_case identifier rather than three words. Code mode is what supplies that missing intent.

yazses features enable code
yazses restart

Activate it with the dedicated command key while you dictate. Spoken symbols become punctuation, and word groups become cased identifiers (snake, camel or pascal).

yazses features info code   # exact behaviour and an example

Maths and LaTeX

If you write papers rather than programs, spoken maths is injected as LaTeX:

yazses features enable math

Saying "x squared plus y squared" injects x^{2} + y^{2}. Common expressions work with no extra dependency; deeply nested expressions need the optional mathspeech extra.

Driving git by voice

yazses features enable gitvoice

A structured grammar maps spoken phrases onto git operations. The important part is the safety design: destructive operations wait for a spoken confirmation, and the undo is always spoken back to you. Say "force push" and it does not immediately force-push — it asks, and it tells you how to reverse it.

Terminal safety

Dictating into a shell is the highest-risk place a misrecognition can land. The terminal command safety gate exists for exactly that:

yazses features enable cmdsafety

It intercepts commands that would be destructive before they run, rather than trusting that the transcription was correct.

The text-target guard is on by default

If no editable text field has focus, YazSes does not type your words somewhere arbitrary. The transcript is copied to the clipboard instead and the tray icon turns yellow. This prevents the classic dictation accident of a paragraph being interpreted as keystrokes by whatever window happened to be in front.

Editor navigation and commands

Beyond text entry, a fast regex command grammar maps spoken phrases to real key sequences — "undo that", "save file", "go to line 42". Several optional capabilities extend this for development work:

Capability Enable with What it does
Voice Jump-to-Symbol yazses features enable jump Jump to a symbol by name
Voice Fuzzy File Open yazses features enable fileopen Open a file by speaking part of its name
Spoken Shell Pipeline Builder yazses features enable shellpipe Compose a shell pipeline by voice
Spoken Regex Builder yazses features enable spokenregex Describe a regex in words
Auto-Pairing & Wrap yazses features enable autopair Balance brackets and quotes
Voice Case Transform yazses features enable casetransform Re-case an identifier by voice

Browse all 139 with yazses features, and see what any one of them does with yazses features info <name>.

Editor context

YazSes can prime the recogniser with context from your editor's language server, so identifiers already present in the file you are working in are more likely to be transcribed correctly. It is off by default:

[commands]
lsp_enabled = true
lsp_editor = "neovim"

You can also maintain a personal vocabulary of project-specific names the model keeps mishearing.

Honest limits

If voice coding is your primary interface, look hard at Talon. Talon is built around exactly this — a deep, scriptable voice-coding ecosystem with a large community grammar library. YazSes is dictation-first, with code mode and commands as strong built-in extras rather than a scripting platform. See the comparison; the two can coexist.

YazSes is the better pick when you want code dictation that works out of the box, is fully open-source, and shares one offline install with your prose dictation, meeting capture and transcription.