Voice transcription for IDE extension

I use superwhisper.. its perfect for dictation

This would be super helpful!

Would love this. I've jumped through hoops today trying to find a way to make this work, but at present, VScode only seems to allow dictation in a file, or with the other built-in chat.

Just use Superwhisper.

@anfedoro not everyone is on a Mac! A solution for WSL would be very welcome.

This feature will really help to enhances the experience.

Up !!!

Can we have both dictation and a voice back for a proper conversation?

Another upvote here... Being able to work on the cli alone and having a /speak command to be able to input a prompt without the need of any external tool would be great. Specially for us in Linux looking for a minimal setup, and not having to used 5 tools in tandem.

@anfedoro not everyone is on a Mac! A solution for WSL would be very welcome.

Well.. not a rocket science to run your own local Whisper model on Linux or Win, and be fully independant. I won't that such voice feature consume my Codex tockens limits... when I can have it for free.

The Codex App supports this:

Use your voice to prompt Codex. Hold =Ctrl+M= while the composer is visible and start talking. Your voice will be transcribed. Edit the transcribed prompt or hit send to have Codex start work.

It'd be great if this can be added to the CLI, too.

Yes, I saw that app feature. If it comes to WSL I will certainly give it a try.

VSCode Codex extension voice input when?
Google Antigravity already has the feature.

https://github.com/openai/codex/pull/3381

VSCode Codex extension voice input when? Google Antigravity already has the feature.

+1

Voice dictation support in available in the Codex app. Implementing it in the Codex IDE Extension is much tougher (it's not clear it's possible) because of the limitations imposed by VS Code's extension mechanism. Google Antigravity is a fork of VS Code, so they don't have the same limitations. We have more flexibility with the CLI (TUI).

+1 from a Linux user. Voice dictation / push-to-talk in Codex CLI is a high-priority workflow need.

@etraut-openai I can't use voice dictation feature in Windows Terminal when running pwsh (not WSL). Is this feature currently supported?

@doggy8088, this feature is under development and is not yet supported.

Thought the voice mode for cli was already supported in https://github.com/openai/codex/pull/3381, just need to enable it in config.toml:

[features]
voice_transcription = true

Thought the voice mode for cli was already supported in #3381, just need to enable it in config.toml: `` [features] voice_transcription = true ``

This doesn't seem to work with codex v0.111.0 and features.voice_transcription = true (which produces the correct experimental warning).

Holding spacebar just inserts a bunch of spaces.

Good news: the Windows / WSL Codex app is out and it supports voice!

+1 (vscode extension)
@etraut-openai , @nornagon-openai , @seratch

Voice dictation support in available in the Codex app. Implementing it in the Codex IDE Extension is much tougher (it's not clear it's possible) because of the limitations imposed by VS Code's extension mechanism. Google Antigravity is a fork of VS Code, so they don't have the same limitations.

If I understood correctly, the main difficulty is that a VS Code extension cannot easily access the microphone due to extension sandbox limitations.

If so, perhaps one possible approach would be to run a small local speech daemon that handles:

• microphone access
• (optional) Voice Activity Detection
• speech recognition

The VS Code extension would then only handle:

• push-to-talk
• UI
• sending START/STOP commands
• receiving the transcribed text
• (optional) receiving the sparse mic level (for UI)

Communication could be done via stdio or a local socket.

This would avoid needing microphone access inside the extension itself.

flowchart LR

User["User (Push-to-talk)"]
Ext["VS Code Extension"]
Daemon["Speech Daemon"]
Mic["Microphone"]
ASR["Speech Recognition"]

User --> Ext
Ext -->|"START / STOP"| Daemon
Mic --> Daemon
Daemon --> ASR
ASR --> Daemon
Daemon -->|"Transcript"| Ext

classDef highlight fill:#e3f2fd,stroke:#1e88e5,stroke-width:2px;
class Ext,Daemon highlight;

<p align="center">
Possible architecture using a local speech daemon</p>

This would be really great to have in the extension as well. Looking forward to it if it ever becomes feasible.
Thanks for the great work, as always!

This doesn't seem to work with codex v0.111.0 and features.voice_transcription = true (which produces the correct experimental warning).

Holding spacebar just inserts a bunch of spaces.

Same on Linux. Why is it so difficult to connect to Pipewire?

Claude Code just released /voice for input dictation

we need this back please for the CLI!

@etraut-openai Is there an estimated release date? I am using ChatGPT Pro on Windows to log into the Codex CLI and still cannot use voice input, but it is already available on macOS.

I think this macos foxuse it's wrong. You should support Linux better.

<img width="589" height="491" alt="Image" src="https://github.com/user-attachments/assets/eb5b1fb0-141f-4b64-87df-afa436eb97e5" />

Claude Code IDE extension for VS Code seems to have this feature already as well and some users in openai's dev forum have been asking for it since Dec 2025

what's the holdup?

also need this because look like already have on codex app

I have created my own "Super Whisper" for Linux: https://github.com/phplego/mywhisper
Maybe someone will find it useful.

Anyway, it would be really useful and straightforward if the extension had its own voice input.

+1 for push-to-talk dictation in Codex, especially if the transcript lands in the prompt box for review before it is sent.

A few edge cases that matter a lot for non-English AI coding workflows:

• mixed-language prompts are common: Chinese users often say things like “帮我 refactor 这个 React hook” rather than pure Chinese or pure English;
• the user should be able to edit the transcript before Codex executes anything, because ASR mistakes can easily become wrong tool instructions;
• language hints and glossary/hotword hints would help with filenames, APIs, framework names, and repo-specific terms;
• voice transcripts should be treated as sensitive prompt material: avoid logging raw audio paths, raw transcripts, cleaned transcripts, or final user-confirmed prompts by default;
• if local/on-device transcription is supported, fallback to cloud transcription should be explicit rather than silent.

For reference, I’m maintaining an open-source local-first Chinese / mixed Chinese-English voice typing project for macOS: https://github.com/botaruibo/MyVoiceTyping

Different product surface from Codex, but the AI coding prompt workflow is very similar: hold to speak → clean up the text → insert into the active input → user reviews before sending.

I'm on Windows so I tend to use the Windows Key + H for native voice to text, but tbh it's not great, doesn't automatically do punctuation, and keeps changing because to 'cause which I find very annoying. The only benefit is that it's built in and I didn't waste any time setting something up.

I don't understand why OpenAI haven't added this yet, they clearly have had good voice capability for a while. Now the Claude vscode plugin's voice is really very good as others have already pointed out above. This isn't a good look for codex in vscode...