OpenAI is developing a new bidirectional voice capability for ChatGPT that will allow the AI to listen and respond simultaneously, according to code strings discovered in the latest app builds and corroborated by early user testers. The feature, tentatively slated for initial testing in June 2026, marks a radical departure from the app’s current push-to-talk and voice activity detection modes, which only permit one speaker at a time.
If successful, the upgrade would close one of the most glaring gaps between human conversation and AI interaction: the ability to interrupt, interject, or speak over one another naturally. Early demonstrations, glimpsed through app teardowns and user-recorded sessions, show the assistant continuing to speak even as a user starts talking—then seamlessly adjusting its response in real time.
The Half-Duplex Handcuff
ChatGPT’s Advanced Voice Mode, rolled out widely in 2024, was a leap forward in realism—offering expressive tones, emotional inflections, and near-instant responses. But beneath the surface, it operated in half-duplex: when the assistant spoke, the microphone effectively shut off; when the user spoke, the model waited until silence before generating a reply.
This technical constraint created stilted turn-taking. Users complained about awkward pauses and the inability to cut the AI off mid-sentence without tapping a button. A bidirectional, full-duplex system would keep the audio input channel open at all times, allowing overlapping speech. The AI would hear user interruptions, process them, and decide whether to halt, pivot, or continue—just as a human would.
Code Trails and Early Sightings
Evidence for the feature first surfaced in late 2025 through “app teardowns” by security researchers and Android developers who decompile APKs for unreleased code. Strings like bidirectional_voice_enabled, <allow_overlap>, and duplex_mode appeared in the ChatGPT Android app’s resources, alongside parameters for sensitivity thresholds and interruption handling.
By early 2026, a handful of users in a closed beta reportedly gained access to a toggle labeled “Duplex Conversations” under Experimental Settings. Leaked screenshots posted to tech forums showed a description: “Allow ChatGPT to listen while speaking so you can interrupt naturally.” The toggle was initially non-functional, but later updates activated it for a small test group.
One participant, speaking on condition of anonymity, described the experience as “uncanny. I started talking while it was explaining something, and it just… paused, acknowledged my interjection, and then gracefully returned to its original point. No lag, no glitch.”
How the Demos Work
Recordings from these early demos reveal a sophisticated processing pipeline. When the user begins speaking over the AI, the model almost instantly stops generating new tokens for its own voice, processes the incoming audio, and decides on a reaction. In some cases, it says something like “Oh, good point—” before adapting its reply. In others, it seamlessly incorporates the user’s interjection into its ongoing sentence.
Latency appears to be the key engineering challenge. The system must transcribe the user’s words, run them through the language model, and decide whether to interrupt output—all in fractions of a second. Leaked documentation hints at a specialized “interruption classifier” that wears the dual hat of voice activity detection and intent analysis, distinguishing between a meaningful cut-in and background noise.
Crucially, the demos show that the microphone does not revert to a push-to-talk paradigm. It stays active throughout the conversation, but OpenAI explicitly states in UI mockups that audio is processed on-device until an utterance is deemed an interruption, after which encrypted audio is sent to the cloud. This split architecture aims to address privacy worries about perpetually open mics.
Timeline: June 2026 and Beyond
Multiple code references include the timestamp string 2026-06-01T00:00:00Z as a default activation date, fueling speculation that a broader test—perhaps for premium subscribers—will start in June. OpenAI has a track record of staging feature rollouts through its ChatGPT Plus and Pro tiers before free-tier availability, so a similar path is expected.
Though June is cited as the start of testing, a full public release could take months. Internal quality targets reference metrics like “interruption false-alarm rate below 2%” and “end-to-end response time increase under 150ms.” Those thresholds must be met across diverse network conditions and languages before the feature graduates from beta.
Why Windows Users Should Care
While the earliest tests will target Android and iOS, the implications for Windows are direct. OpenAI’s desktop app for Windows already supports Advanced Voice Mode via a system tray toggle and hardware mic. A bidirectional upgrade would naturally extend to that environment, potentially allowing users to hold fluid conversations while working in other windows—no need to click a microphone icon.
Moreover, Microsoft’s deep partnership with OpenAI means that any ChatGPT innovation often spills into Windows Copilot and other AI-powered tools. A full-duplex voice model could eventually underpin voice interactions across the operating system, from File Explorer searches to real-time meeting summaries in Teams. Windows insiders have already spotted references to a “Copilot voice duplex” flag in early 2026 builds, though Microsoft has yet to comment.
For PC users, bidirectional voice would also pair well with the growing array of Copilot+ PCs sporting neural processing units (NPUs). On-device interruption detection could leverage NPUs to minimize cloud dependency and latency, a design choice that aligns with Microsoft’s emphasis on local AI processing for privacy and speed.
The Competitive Race to Full-Duplex
OpenAI isn’t alone. Google’s Gemini Live voice mode, launched in 2024, allowed interruptions but still operated in a half-duplex rhythm: the assistant stopped speaking the moment it detected voice, but re-processing lag made it feel less seamless. Apple’s Siri, even after its 2025 revamp, remains largely half-duplex. Amazon’s Alexa has long enabled “barge-in” on smart speakers, but the underlying architecture is simpler—it doesn’t require the deep contextual understanding a large language model provides.
ChatGPT’s approach differs because it aims to maintain conversational coherence even when interrupted. In demos, the assistant often acknowledges the interruption explicitly, then weaves its response into the revised context. This level of nuance demands an advanced model that can simultaneously process two overlapping semantic streams—what researchers call “co-speaking.”
A leaked internal OpenAI benchmark, titled “InterruptQA,” evaluates the assistant on its ability to correctly answer a question after being interrupted mid-sentence with a clarification, contradiction, or topic switch. Early results reportedly show the duplex model outperforming a baseline non-interruptible system by 23% on answer accuracy and 41% on user satisfaction scores.
Technical Hurdles and Acoustical Challenges
Achieving true full-duplex voice on everyday devices is non-trivial. Echo cancellation must be perfect to prevent the AI from misinterpreting its own audio output as a user interruption. Background noise, room reverberation, and varying microphone quality add complexity. The leaked code includes a pre-processing block labeled “neural AEC” (acoustic echo cancellation) that appears to leverage machine learning to filter out the assistant’s voice before interruption detection.
Another hurdle is model bias. If the system becomes too sensitive, a cough or a dog bark could derail the conversation. Too insensitive, and the user is shouting over the AI to get a word in. The sensitivity slider seen in beta screenshots—ranging from “Polite” to “Assertive”—hints at OpenAI’s solution: users can tune how aggressively the assistant yields the floor.
Network jitter also plays a role. Because the assistant must receive the user’s audio, process it, and decide to stop its own streaming speech, any delay can result in awkward overlaps or “clipped” words. OpenAI’s technique, according to an engineering blog post from 2025, involves speculative generation: the assistant continues speaking while simultaneously computing potential responses to a likely interruption, ready to cut over instantly.
Privacy: Always Listening, but Not Always Streaming
The concept of an always-active microphone makes privacy advocates nervous. OpenAI’s documentation for the experimental feature stresses that audio is processed in a privacy-preserving loop. Voice activity detection and the initial interruption gating run entirely on-device, using a lightweight model that can distinguish between ambient noise and speech. Only when the system classifies an utterance as a genuine interruption (with high confidence) does it send the relevant audio clip to OpenAI’s servers for full processing.
Users can also disable duplex mode entirely, fall back to push-to-talk, or review and delete voice transcripts. Windows users would benefit from built-in privacy indicators—like the taskbar microphone icon—that already light up when any app accesses the mic. OpenAI’s settings page includes an option to require a physical button press before the assistant enters listening mode, though that somewhat defeats the purpose of bidirectional conversation.
What Early Testers Are Saying
Feedback from the limited beta has been largely positive but spotted with caution. One user on a popular Windows forum wrote: “It’s like talking to a person who’s actually paying attention, not a robot waiting its turn. But it does get confused if three people are speaking at once.” Another noted: “I interrupted to ask for a pizza topping recommendation while it was listing Sci-Fi movies, and it pivoted mid-sentence without missing a beat. Spooky.”
Critics point to occasional “backtalk” where the assistant attempts to finish its thought while the user is already two sentences ahead, resulting in a jumbled mix. These edge cases, OpenAI engineers reportedly believe, will smooth out as the interruption classifier learns from millions of real-world interactions.
The Bigger Picture: Human-Like AI Conversations
Bidirectional voice isn’t merely a convenience feature—it’s a cornerstone of natural human communication. Linguists note that overlapping speech accounts for up to 20% of everyday dialogue, serving as backchanneling cues (“uh-huh,” “right”) or collaborative turn-taking. Without the ability to overlap, AI conversations feel rigid and transactional.
OpenAI’s push toward full-duplex aligns with its broader mission to make AI assistants more present and useful. A recent company blog post hinted at “anthropic voice patterns” being a core research area, though it stopped short of announcing features. Industry observers see bidirectional voice as a prerequisite for always-on ambient assistants that participate in meetings, brainstorming sessions, or even casual banter around the house.
What’s Next for Windows Enthusiasts
For those running Windows 11 or the forthcoming Windows 12, the June 2026 test window offers a glimpse of a more integrated AI future. If the feature lands first on mobile, expect the Windows ChatGPT app to follow within weeks, given OpenAI’s unified codebase strategy. Users can prepare by ensuring their microphone setup is solid—headsets with good noise cancellation will likely deliver the best experience.
In the meantime, the leaked code and user reports serve as a roadmap. They suggest that conversations with AI are about to become dramatically more fluid, and the line between human and machine dialogue will blur further. As one developer who analyzed the APK put it: “It’s the last piece of the puzzle for a truly conversational assistant. Once this ships, you’ll forget you’re talking to code.”
OpenAI declined to comment on “unreleased features,” but with June 2026 on the horizon, the countdown to the end of half-duplex voice has begun.