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

+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 (#72284, #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.

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.

Expanding on and adding to the use cases described here.

I'm on Max, running Cowork on Linux desktop for design and architecture work.

Cowork currently has dictation but not voice mode. This is confirmed
in the support docs — "while dictation is available in Claude Cowork
and Code, voice mode is not"
(https://support.claude.com/en/articles/11101966-use-voice-mode).

The workflow this blocks: I'm reading a specification or design
documentation on screen while directing Cowork through implementing
against it. My eyes are on the documentation or the code itself while iterating along side Claude.
Dictation solves the input half, but I still have to
break away and read Claude's responses — which puts my attention
right back where it started. The value isn't Just hands-free input on its
own; it's spoken output that lets me stay on the source material
while the agent works.

This overlaps with #58429 and #64347, but the driver is different.
Those are framed around accessibility. Mine is about sustaining
attention on external material while directing an agent — reviewing
a doc, reading a paper, working through a spec, Reviewing/Writing Code.
Same feature, and I suspect a much larger set of users than the framing so far suggests.

Voice mode already exists in chat. The gap is that it isn't wired into Cowork, where the long-running work actually happens.