Turn-taking was a training wheel. Full-duplex removes it.

OpenAI's GPT-Live-class models listen while they speak: barge-in, mid-sentence pauses, brief acknowledgments, fewer robotic handoffs. That is a real interface change for consumer chat. It is also a product design problem. Solo builders who bolt "voice mode" onto a text workflow without redesigning state will ship a charming demo that cannot be supported.

What full-duplex actually changes

In turn-based voice, the state machine is simple: user speaks, system thinks, system speaks, repeat. In full-duplex, events overlap. The user interrupts. The model hedges. A tool call starts while audio is still playing. Latency hides inside conversation.

That means your product needs:

  • An explicit conversation state (listening / speaking / tool / waiting-for-confirm)
  • Interrupt policy (what cancels, what queues)
  • A transcript that remains the system of record
  • Safety rails that still fire when speech is messy

If you only stream audio and hope, you will not be able to debug a single bad session.

When voice is the right interface

Voice wins when hands or eyes are busy and the task is conversational:

  • Field notes while moving
  • Coaching / practice with immediate feedback
  • Accessibility paths where typing is costly
  • Quick capture that will be edited later in text

Voice loses when the task is precise, auditable, or multi-hour:

  • Code review you must defend
  • Legal or medical wording
  • Anything that needs a diff
  • Anything irreversible without a second factor

A good rule: if you would not accept the outcome without reading it, do not make voice the only channel.

Pattern library for solo products

Barge-in: User interrupts → cancel speech, keep partial intent if confident, else ask a short clarification.

Silence: Distinguish thinking silence from dead air. Show a light listening state. Do not narrate every internal step.

Tool handoff: When a tool runs, say one sentence, then go quiet. Dump raw tool JSON into the transcript, not into the ear.

Confirm irreversible: Switch to explicit confirm ("say publish to publish"). Free-form "yeah" is not enough for money or delete.

Fallback to text: Always offer a session log and a continue-in-text path. Voice is a mode, not a prison.

The API lag trap

Consumer apps ship experiences that developer APIs may not match yet. Full-duplex in a chat app does not automatically mean full-duplex in your backend. Design your product against the API you can actually call, not the keynote.

If your architecture needs server-side tool use with enterprise auth, verify that path before you market "talk to your stack."

The catch

Natural conversation hides errors. Users trust a calm voice more than a correct system. That is dangerous. Keep the transcript authoritative. Log model decisions. Let humans reconstruct what happened.

Safety regressions matter more in voice because sessions move faster and social pressure to "just continue" is higher. High-risk topics need the same policies you would enforce in text — preferably stricter.

Two product sketches

Good fit: A language practice app. Full-duplex lets learners interrupt. Transcript becomes homework. No irreversible tools.

Bad fit: A deploy console. Voice can query status. Voice must not production-deploy without a typed confirm in a separate channel.

Bottom line

Full-duplex is a better microphone, not a better architecture. Use it where conversation is the job. Keep text and audit where correctness is the job. Solo builders who respect that split will ship voice that feels alive without becoming unsupportable.