Dictate into the composer instead of typing. The transcript lands in your draft for review, and only sending the message sends anything to an agent. Dictation also works in citation comments, diff and file annotations, pull request comments, replies, and review summaries, and in the desktop app's browser annotation editor.
On iPhone and Android, voice input transcribes on the device by default. To opt in to a named environment on that phone, open Settings → Voice → Transcribe on. See Voice input on mobile.
On web and desktop, you choose where transcription runs. Nothing is sent until you choose one: open Settings → Voice and set Transcribe with to the environment that should transcribe for you. On a new device, the microphone or dictation shortcut opens that setting from the composer. Microphone audio recorded on this device is sent to that environment over your authenticated connection, transcribed there, and discarded. The choice is saved on this device only, so every computer and browser you use asks for it once. Set it back to Off at any time; a recording in progress is discarded.
If the chosen environment goes offline, voice input waits for it. T3 Code never substitutes another environment.
For cloud dictation, choose OpenAI or Grok Voice Transcribe via the environment you want to use. If it needs a key, enter one and choose Save and enable. You can also choose a secret provider instead of Value and enter a reference. Unlock the vault on that environment before dictating. Anyone with dictation access to the environment can use the key, but only this device opts in. If a key is already saved, just confirm the provider. After you stop, the environment forwards your recording and any vocabulary hints to the selected provider, which bills its API account. No local model download or chat subscription is needed. Switch back to This device on mobile or Off on web and desktop to stop sharing. Removing the key stops that provider for everyone on the environment.
OpenRouter works the same way but has no default model: pick one from its live speech-to-text catalog when you enable it, or later under Manage OpenRouter. The model choice is yours alone and needs no settings access, so anyone with dictation access can pick a different model without touching the saved key. OpenRouter chooses who runs the model. T3 requests zero data retention; your OpenRouter account can also allow safety retention. If no endpoint qualifies, dictation fails instead of switching models. Vocabulary hints are not sent to OpenRouter; dictionary spellings are applied after transcription. If OpenRouter stops listing your model, dictation stops until you choose another.
In Settings → Voice → Transcription, download a model and choose Use model if it isn’t already selected. Test model lets you try it without sending a message. If you use the composer microphone before a model is installed, it offers to download one. The recommendation uses your Voice language, or your system’s preferred languages with Auto-detect.
Select the languages you dictate in to compare models that support all of them. Accuracy shows how often the model got words or characters wrong in each language; speed is the time to transcribe 30 seconds on a reference GPU. Choose All languages to browse the full catalog. Filtering does not switch your model.
Advanced controls GPU or CPU use and when an idle model is unloaded. With hardware set to Auto, a recording that fails on the GPU is retried once on that same environment's CPU; a GPU you picked yourself reports the error instead. Nothing is ever retried on another environment. Models marked Live preview show text while you speak. Other models can process long recordings in chunks at pauses while you record; models with built-in long-form support transcribe when you stop.
Press the microphone in the composer, or mod+shift+d (mod is Command on
macOS and Ctrl elsewhere). Press again, or the check mark, to finish; Esc
discards the recording and leaves your draft unchanged. Recordings last up to
five minutes. Text is inserted at the cursor or replaces the selection you had
when recording started, and the editor is read-only while dictation runs.
Pressing Send while dictating finishes the recording and sends the draft once the transcript has been inserted. Cancelling also cancels that queued send. In the other editors, Save, Comment, Reply, Submit review, or Attach works the same way, and the shortcut targets the editor that has focus, or the composer when none does. In the desktop browser's annotation editor, the microphone and the shortcut work while the page has focus; the page you are annotating never sees the recording, which runs in T3 Code itself.
Dictation shortcut in Voice settings picks how the keybinding behaves: Auto tap-toggles or records while held, Hold records only while the keys are down, and Toggle always needs a second press. Change the chord itself in Keybindings settings.
- Language picks what you speak. Auto lets a multilingual model detect it. Models that translate offer Translate to English.
- Dictionary holds names and terms the model should get right. Add an alias under a term when transcription keeps producing a particular misspelling. Your dictionary is saved on this device. The name of the project you dictate in is sent along as a hint.
- Project dictionary is vocabulary saved with a project on its environment, so everyone who dictates in that project gets the same spellings. Pick the project at the top of Settings to edit it; with an environment selected, the entries apply to every project there. Project spellings win when they overlap your personal dictionary. Each recording carries up to 100 entries from the two combined.
- Silence filtering is on by default. Your chosen environment skips non-speech audio before transcription. Turn it off to keep those sections. Filtering happens after upload: it does not grant permission to send audio or reduce how much is sent.
- Remove filler words drops hesitation sounds from finished transcripts.
Language and these preferences travel with each recording to the chosen environment; nothing else about the project is sent.
Transcript cleanup is a separate choice and is off by default. When on, the finished transcript and up to 1,024 characters each from before, inside, and after your selection are sent to the text generation provider configured on the draft's environment, which fixes punctuation, numbers, and spoken corrections such as "make it 42, sorry, 24". Depending on the provider, that can involve an external service. Audio is never sent to it. Use the built-in instructions or write your own, press Skip while cleanup runs to insert the raw transcript, and if cleanup fails the raw transcript is inserted.
Cleanup uses the environment's text generation model unless you pick a Cleanup model for that environment, for example a cheaper or faster one.
If cleanup misses the way you correct yourself, set a Correction cue: the word or phrase you say to flag a correction, such as "scratch that". Cleanup applies the correction only where the context shows one and drops the cue, and the model hears the cue as a dictionary hint.
In Settings → Voice, Test microphone beside the microphone selector records up to ten seconds and plays it back in your browser; the audio stays on this device and is never transcribed, so silent playback points at the microphone. Test model on the selected model shows a transcript using the same environment, language, and dictionary as the composer, but skips transcript cleanup even when it is on, so you see what the model heard. Browser recording needs HTTPS or localhost and microphone permission; for a remote server over plain HTTP, use the desktop app. If the selected microphone is missing, choose another or the system default.