Transcribe dictation while recording to reduce post-stop delay

What variant of Codex are you using?

App (desktop)

What feature would you like to see?

Start transcribing composer dictation while the microphone is still recording.

Today, the app appears to wait until the user clicks Done before sending the completed recording for transcription. This leaves the full transcription step until after recording and can make the wait grow with longer dictation.

Please stream microphone audio to one transcription session throughout the recording. When the user clicks Done, send the remaining audio, finalize the session, and insert one editable transcript into the composer.

This does not require showing interim text while the user speaks. The app can continue displaying only the final transcript.

Expected behavior:

• Start processing after a short initial buffer.
• Keep one stateful transcription session so context is preserved across audio chunks.
• On Done, flush the remaining audio and wait for the final transcript.
• Keep the recording available for a batch retry if streaming fails.

Additional information

This is separate from #23012, which requests visible interim text while recording.

#36803 documents 3 to 17 seconds of post-stop transcription latency and mentions a streaming path, but it does not require transcription to begin during recording.

The public app-server API already accepts live audio through thread/realtime/appendAudio and emits transcript delta and done events. However, thread/realtime/stop currently closes the realtime session. Dictation may need an explicit flush or commit step so the last audio is transcribed before the session closes.

I could not find the Desktop composer implementation in this repository, so some of the client work may live elsewhere. I would be happy to contribute a focused change if the team agrees with the approach and invites it under the repository contribution policy.

Start by reading the app-server realtime flow at thread/realtime/appendAudio and thread/realtime/stop, then locate the Desktop composer implementation, which the report says may be elsewhere. Confirm how a stateful session can flush remaining audio before closing and how recording is retained for batch retry. Done means transcription begins during recording, finalizes after Done, and inserts one editable transcript.