Repository navigation
[Feature]: Built-in voice dictation for the Linux desktop app #30262
Description
Request
Add voice dictation to the composer in the Linux (and Windows) desktop app, matching the macOS dictation that shipped in #24938, and include a built-in local transcription engine so users don't need a separate dictation tool.
Problem
Right now the desktop composer on Linux has no mic control. The only way to dictate is a separate system-wide tool, such as Handy or nerd-dictation, that types into the focused window. That has real drawbacks:
- Many users don't have such a tool and won't install one just for this.
- On GNOME Wayland, the usual way of typing into another window (
wtype) doesn't work, so these tools fall back to the clipboard orydotool, which is fragile. - macOS gets OS speech APIs, but Linux has no system speech API, so T3 Code can't rely on the OS the way the macOS build does.
Desired behavior
Same UX as #24938:
- A mic button in the composer, plus a keyboard shortcut (push-to-talk or toggle).
- The transcript goes into the draft as editable text and is never sent automatically.
- Recording and transcribing state is visible, with a cancel action.
- Clear errors for microphone permission, missing input device, and failed transcription.
Built-in engine
- Ship an on-device engine that works offline, such as whisper.cpp or Parakeet. The model would be a one-time optional download, picked by size or quality in settings.
- Reuse the engine picker and bring-your-own-key options from [Feature]: Choose transcription engine (BYOK — Gemini/OpenAI/Groq) for macOS voice dictation #28149 (Gemini, OpenAI, Groq) as alternatives, and the vocabulary hints from Voice input: project-aware vocabulary hints for speech-to-text #27752.
- State clearly whether audio leaves the machine for each engine.
Platform
Linux desktop (AppImage, nightly channel), tested on Ubuntu with GNOME on Wayland. Windows desktop likely has the same gap.
Reacted by Joris Grau, Augustya Baghel, Osman Dağdeviren, Daniil, Jens Martsch and Felix Ostmann
Activity
JorGra commentedon Nov 3, 2026
This is a must have for me as well T3 Code looks promising, but i need voice mode and other solutions are not great, maybe there is a way to use the voice mode Anthropic provides through the T3 Code interface?
vibhor1102 commentedon Nov 4, 2026
I’d like built-in dictation in the Linux desktop composer too. ChatGPT/Codex’s dictation is a useful part of the experience: record speech, receive editable prompt text, review it, and send when ready.
Please also investigate using Codex’s existing transcription infrastructure. Its open-source backend includes an experimental transcription-only realtime mode. Whether that can use existing ChatGPT authentication needs verification; the WebSocket path inspected currently requires an API key.
My request is speech-to-text in the composer, with a microphone control, shortcut, cancellation, and review before sending. Full voice conversations aren’t necessary for this.
Environment: T3 Desktop
0.1.8-nightly.20261104.4165, Linux/KDE Plasma.Drafted with Codex (gpt-6.1-sol) in T3 Code on my behalf.
Sadrak commentedon Nov 5, 2026
Same.
Best would be a fast on the fly visible transcript for parallel reading of complex words / terms. To stop recording, fixing / adding words to the glossar, and continuing recording.
vibhor1102 commentedon Nov 5, 2026
Live transcript, althought convenient, has inherent tradroff of weaker context-aware autocorrect that comes from processing your output after you've completed the full dictation. I'd happily take a stronger post-dictation processing over more mistake-prone live transcription on any day.