Feature request: continuous voice conversation mode (like mobile app) in Desktop/Cowork

[kenta4es]

Please bring the mobile app's live voice mode to Claude Desktop (Cowork).

Desired behavior:

• Always-listening mic toggle, streaming STT + TTS duplex (no send button)
• Conversation continues WHILE the agent executes tasks in background (subagents/detached jobs)
• Agent can speak status updates mid-task

Use case: hands-free orchestration of long-running local pipelines (trading research: Python + n8n + API jobs). Today the desktop flow is strictly request-response with two dead pauses (user typing / agent silent while working). The mobile app already has live voice — feature parity requested.

Posted on behalf of a Cowork research-preview user (Windows).

[david-tapias]

+1 from another daily Cowork user (Windows 10 x64, v1.22209.3) — adding a different angle: non-English speech quality and accessibility.

I'm a 57-year-old Spanish/Catalan speaker and Cowork is my main work tool several hours a day. My experience with the current options:

• The built-in mic dictation (Cowork's own): it transcribes into the input box, but it's not a conversation — and in Spanish it degrades my speech before Claude sees it. Real example from my sessions: I said "Wallapop" (a Spanish marketplace) and the dictation wrote "Guadalupe". Claude untangles part of it, but a lot of intent dies at the transcription layer — the only layer Claude cannot fix. There are also open reliability reports for this mic on Windows ([BUG] Cowork microphone input cuts off after ~2 seconds on x64 — works fine on ARM64 #72284, [BUG] Cowork voice dictation: mic activates briefly on 2nd attempt, then stops (Windows 11) #72469).
• The mobile app's voice mode: understands me perfectly — but it can't see my desktop workspace, files or browser, so it doesn't help for real work sessions.

What I'd add to the original request:

1. Speech understood directly by Claude (or an STT that handles Spanish, Catalan and code-switching well) — today the dictation layer is where my intent gets lost.
2. Accessibility: typing for hours is tiring at 57; voice would make long working days sustainable — same for anyone with hand strain or motor difficulties.

Happy to test any beta and provide Spanish/Catalan examples.

[creativityfromthefuture]

Adding a third use case, from a Cowork research preview user on Windows.

My workflow is hands-busy but eyes-on: I run long diagnostics and multi-step jobs across three monitors while the agent works. I'm reading the screen, not the keyboard. The current dictation flow is click mic, speak, click stop, press Enter - four actions per turn, and every one of them pulls my hands and attention back to the keyboard. Over a long session that's the dominant friction, more than model latency.

Two asks, and I want to be explicit that the first is much smaller than the second:

Minimum that would solve most of this: always-listening dictation with VAD auto-submit in Cowork. No TTS, no duplex audio, no barge-in - just a mic toggle that stays on and submits on a natural pause, so a turn is "speak" instead of "click, speak, click, click." This seems like a much smaller change than full voice mode and would recover most of the value.

Ideal, as originally filed: full duplex with streaming TTS, plus the agent speaking status updates while it works. That second half matters more than it sounds - right now there's a dead pause while the agent executes and no way to know what's happening without watching the transcript.

On why Cowork specifically: voice mode already exists on Desktop and mobile, but per the docs it can't reach Cowork projects, skills, or file and screen access. So the surfaces that have voice can't do my work, and the surface that does my work has no voice. Switching apps doesn't solve it.

Agreeing with david-tapias on the reliability bugs (issues 72284 and 72469) - always-listening only helps if the mic stays alive, so those may need to land first.