# OpenAI's full-duplex voice mode on the ChatGPT desktop app: what holds up and what doesn't > Full-duplex voice on the ChatGPT desktop app nails interruptions but chokes on noise and structured output — here's where it fit my workflow and where it didn't. **Source:** Broke It (https://brokeit.dev/posts/openai-s-full-duplex-voice-mode-on-the-chatgpt-desktop-app-what-holds-up-and-wha/) **Published:** 2026-07-25 **Original source:** https://techcrunch.com/2026/07/24/openais-new-voice-mode-makes-it-to-the-chatgpt-desktop-app/ **Topics:** product-launch, voice, assistant --- Full-duplex voice on the ChatGPT desktop app works well enough for quick back-and-forth, but it falls apart the moment you need it to shut up and listen. Interruption handling is the good part — you can cut in mid-sentence and it actually stops. Latency is fine on a wired connection and bad on flaky WiFi. I wired it into my daily workflow for a week and here's where it earned a spot and where it got closed. ## Does the interruption handling actually work? Yes, and this is the feature that matters. In the old voice mode you'd talk over the model and it kept going like a podcast on autoplay. In the new full-duplex mode you say "stop" or just start talking and it cuts within maybe 300-500ms. That's the difference between a demo and a tool. According to TechCrunch's report, OpenAI is pushing this as the more natural conversational mode, and on interruption specifically that claim holds. The gotcha: it interrupts *itself* too. Background noise, a cough, a coworker walking by — the model treats ambient sound as a turn boundary and stops talking mid-answer. In an open office it's unusable without a headset. With a decent noise-gating mic it's fine. So budget for hardware, not just software. ## What broke? Latency is inconsistent. On ethernet, first-token audio comes back fast enough to feel like a phone call. On hotel WiFi I saw 2-3 second gaps where I couldn't tell if it heard me, so I repeated myself, which then triggered a double response. There's no visible "I'm listening" state that's reliable enough to trust, so you end up guessing. The other break: it's a conversation tool, not a work tool. Ask it to read back code and it mangles symbols — it'll say "underscore" and "open paren" in a stream that's impossible to follow. Anything structured (JSON, a stack trace, a file path) is better in text. Voice is for thinking out loud and rubber-ducking, not for consuming precise output. ## Where it actually fit my workflow One concrete win: I used it to talk through architecture decisions while pacing away from the keyboard. Describe the problem, let it push back, no typing. That's genuinely useful and I kept it for that. What I did *not* keep it for: anything I needed to copy, anything with exact identifiers, or anything in a shared room. Also worth noting — desktop makes this better than mobile because you've got a real mic setup and you're already at your machine when a follow-up needs to become a text prompt. The handoff from voice to "okay now show me the diff" is one click away. ## Should you use it? Use it if you have a headset and a quiet room and you think better out loud. Skip it if you're in an open office or you expect it to dictate precise output. It's a good conversational front-end and a bad transcription surface. For me it survived the week for one narrow use — talking through problems — and got closed for everything else. That's a real feature, not a gimmick, but it's narrower than the pitch. ## FAQ ### Does the full-duplex voice mode handle interruptions well? Yes. You can cut in mid-response and it stops within roughly 300-500ms, which is the main upgrade over the old mode. The downside is it also stops on background noise, so a noise-gating mic or headset is close to required. ### Is it good for reading back code or structured output? No. It verbalizes symbols like "underscore" and "open paren" in a stream that's hard to follow, and file paths or JSON come out mangled. Keep structured output in text and use voice for discussion. ### Is desktop better than mobile for this? For me, yes. Desktop gives you a real mic setup and a one-click handoff to text when a follow-up needs an exact prompt or you want to see a diff. Mobile lacks that quick transition.