Environment: Claude desktop app, Cowork session, Windows 11. Built-in laptop microphone. Signed in with a Claude.ai account.
Summary: The in-bubble voice dictation (microphone button in the chat input box) transcribes correctly the FIRST time it is used in a session, but fails on every use after that within the same session.
Steps to reproduce:
- Start a brand-new Cowork session.
- Click the microphone in the input box and dictate — works perfectly, for as long as I like.
- Use the microphone a SECOND time in the same session.
- The mic activates and the audio level bars respond to my voice (so audio is being captured), but NO words are transcribed into the box — it shows "..." and then stops after about 3 seconds.
Reproduces in fresh sessions every time (1st use works, 2nd+ fail).
What I ruled out by testing:
- Microphone / hardware / permission: the level bars move when I speak, so capture is fine.
- Network: same connection throughout.
- Account / app build: a brand-new session's first dictation works.
- Session age/weight: it's not old/long sessions — it's specifically the 2nd+ dictation in ANY session, including a fresh one.
Expected: dictation works on every use within a session.
Actual: only the first use per session transcribes; subsequent uses capture audio (level meter active) but produce no text and time out after ~3s. The recogniser appears not to re-initialise on the second invocation.
Impact: high — voice dictation is my primary input method.
Possibly related: #38753, #38712.
Environment: Claude desktop app, Cowork session, Windows 11. Built-in laptop microphone. Signed in with a Claude.ai account.
Summary: The in-bubble voice dictation (microphone button in the chat input box) transcribes correctly the FIRST time it is used in a session, but fails on every use after that within the same session.
Steps to reproduce:
Reproduces in fresh sessions every time (1st use works, 2nd+ fail).
What I ruled out by testing:
Expected: dictation works on every use within a session.
Actual: only the first use per session transcribes; subsequent uses capture audio (level meter active) but produce no text and time out after ~3s. The recogniser appears not to re-initialise on the second invocation.
Impact: high — voice dictation is my primary input method.
Possibly related: #38753, #38712.