Android: microphone dictation into the composer

[widingmarcus-cyber]

Use case

I use T3 Code with a Windows PC and the Android app. I would like to dictate a prompt directly into the Android composer, edit the transcript, and send it when ready.

Requested behavior

• A microphone button in the Android composer.
• Record, stop, and insert the transcript into the current draft.
• Keep the text editable and do not send it automatically.
• Show clear feedback if permission or transcription fails.

The main request is speech-to-text for composing prompts. A full voice conversation is a separate feature.

Existing work

• [Feature]: Mobile voice dictation — mic button in the composer with BYOK OpenAI/Groq transcription #8718 was closed as a feature proposal and directed to Ideas. The closure says Android transcription is not implemented and external-service credentials should live on the environment.
• feat(mobile): environment-backed voice transcription #9028 proposed environment-backed transcription for Android and older iPhones, but was closed without merging.
• feat: Codex voice conversations on web, desktop, and mobile #13168 is an open proposal for broader voice conversations, including Android.

I am open to the supported transcription approach; this request does not require keys to be stored on the phone.

Is Android composer dictation still planned? Is there a current tracking discussion or implementation that replaces #9028, and any intended release milestone?

[OffCrazyFreak]

Hi! Just wanted to add my voice (pun intended 😄) to this one. I'd love to see dictation on Android too. The iPhone version is great, and a mic button in the composer that drops an editable transcript into the draft would be perfect for those of us on Android. Thanks for all the work on T3 Code mobile!

[matterramun]

I've built this for Android and it works on my phone, so I'm putting it here before opening anything. I'd like to know whether maintainers would take it before I send a PR.

Dictation uses the phone's own speech recognizer through expo-speech-recognition. It needs no API key, server endpoint, upload or paid service. The composer already has a mic button; it was hidden on Android because there was no transcriber, and now it shows up. The transcript goes into the draft and stays editable, and nothing is sent automatically. The controller, the button and the error states are the existing shared code, and the iPhone path is untouched. The recognizer is whatever the phone ships with, which on most devices means Google's, so audio may leave the device. I haven't tried to force on-device recognition.

The branch also has a per-thread switch, "Read replies aloud", that reads each finished reply out, which is disabled by default. A button under the thread header pauses and resumes, goes back or forward a sentence, restarts, and reads the last reply again once it has finished. It takes audio focus, so music pauses and comes back afterwards. Together with dictation you can speak, listen and speak again without looking at the screen. That overlaps with #12622, #6762 and #14394. Playback keeps going with the app in the background or the phone locked.

Dictation needed one Android-only fix. The microphone permission dialog pauses the activity, so React Native reports background while dictation is still preparing, and the controller gave up with "Voice input stopped when the app moved to the background." Logcat shows MainActivity pausing behind GrantPermissionsActivity. The controller now ignores background events while the permission request is open. iOS reports inactive there, which is why it never came up.

One gap I want to fix, reading aloud doesn't put a media card in the notification shade or on the lock screen the way other audio apps would, so you can only pause it from inside T3. I'd want to add that before calling the feature complete.

I tested on a Samsung Galaxy (SM-S916U) with a release build against a 0.0.46 nightly server. Unit tests cover the controller fix, transcript handling, the playback queue and sentence splitting. Screenshots are attached. I haven't tried other devices or recognizers, offline recognition, or iOS; the iOS audio-focus code is written but unbuilt.

The code is at https://github.com/matterramun/t3code/tree/feat/mobile-voice-conversation. If it fits, I'd send two PRs: Android dictation first, then reading aloud.

This is my first attempt at contributing here so please let me know if I should change my approach 😅 I built this first for me and discovered there were other folks interested in the feature after the fact, so this is my attempt to share the tokens I've already spent on this.

Screenshots of the current implementation on my device: