Why does an AI that answers your shop's phone suddenly sound like a person, when the same kind of system sounded like a train-station announcement eighteen months ago? The answer is one engineering change with an unglamorous name: full-duplex.
On 10 September 2026, OpenAI released its full-duplex voice model GPT-Live-1 into its public API at US$0.05 per minute for the voice layer, according to OpenAI's announcement. That single move put conversational phone AI inside the budget of an ordinary Hong Kong shop. This guide explains what full-duplex actually means, what it changes, and what it still cannot do.
What is full-duplex voice AI?
Full-duplex voice AI is a voice system that listens and speaks at the same time. It keeps processing incoming audio while its own reply is still playing, so it can stop mid-sentence when a caller interrupts. Half-duplex systems take strict turns: they listen, then speak, and ignore you while speaking.
The word comes from telecommunications, not from AI. A telephone line is full-duplex, because both people can talk at once. A walkie-talkie is half-duplex, because only one side transmits at a time and you have to say "over".
Every AI phone agent before this generation was, in effect, a walkie-talkie wearing a telephone's costume.
How does full-duplex voice AI actually work?
A full-duplex model runs two audio streams in parallel. One stream carries what the caller is saying. The other carries what the model is saying. Both are open continuously, so the model can detect a barge-in within a fraction of a second and cut its own speech.
Older systems ran a relay of three separate steps, each waiting for the one before it:
--- Speech-to-text converted the caller's audio into words, but only after detecting silence.
--- A language model read the text and wrote a reply.
--- Text-to-speech turned that reply into audio and played the whole thing.
That relay produced the two habits everybody hated. First, a pause of one to three seconds before every answer. Second, an agent that kept talking over you because it had already committed to playing the full audio file.
A full-duplex model collapses those steps into one system that hears and talks continuously. In OpenAI's design, the voice layer handles the conversation itself and hands the harder thinking to a separate backend model behind it, which is why the voice layer can be billed separately and cheaply.
What can full-duplex handle that older voice AI could not?
Four specific conversational behaviours separate a full-duplex agent from a half-duplex one. They sound small. In a real phone call they are the whole difference between "acceptable" and "please transfer me to a human".
Interruption handling
The caller says "no, wait, the other branch" while the agent is still describing the first branch. A full-duplex agent stops within a fraction of a second. A half-duplex one finishes its paragraph.
Pause handling
A customer checking a receipt goes quiet for three seconds. Half-duplex systems treat silence as the end of the turn and start talking. Full-duplex systems can model the difference between a thinking pause and a finished sentence.
Backchannelling
These are the small noises humans make to show they are still listening: "mm-hm", "right", "okay okay". A full-duplex agent can produce them, and more importantly can recognise them without treating them as a new question.
Turn-taking
Deciding when to start speaking without cutting the caller off. This is the hardest of the four, and it is the one that makes a call feel either natural or exhausting.
What can a Hong Kong small business use this for?
Voice AI is worth considering wherever a phone rings and nobody can pick it up. In Hong Kong that is a specific and very expensive list, because the busiest hour for calls is usually the busiest hour for service.
The restaurant at 7 PM
A dinner-time booking call arrives while every staff member is carrying plates. The call is missed, and the table goes to whoever answered first. A voice agent takes the booking, confirms the party size, and writes it into the system.
The clinic or salon at closing time
Appointment changes arrive after hours. Each one that goes to voicemail becomes a morning callback, which costs staff time twice.
The shop answering "do you still have it?"
Stock and opening-hours questions are the single most repeated call in retail. They are also the easiest to automate, because the answer is a database lookup.
Outbound confirmations
Reminder calls the day before an appointment reduce no-shows, and almost nobody has staff to spare for them.
A realistic target for a first deployment is one narrow job, not a general receptionist. Bookings only, or stock questions only. Narrow agents succeed; ambitious ones get escalated.
How much does full-duplex voice AI cost to run?
The headline number for the voice layer is US$0.05 per minute, or roughly HK$0.39 a minute at current rates. That is the front-end voice cost only. You pay separately for the backend model that does the reasoning, and for the telephony line that carries the call.
A rough monthly picture for a small shop taking 300 calls a month at three minutes each:
--- 900 minutes of voice at US$0.05 equals US$45, about HK$350.
--- Backend model and tool calls, billed separately, typically add a similar order of magnitude depending on how much the agent has to look up.
--- Telephony and number rental in Hong Kong are a separate line item from a local carrier.
Compare that with a part-time person answering the same calls, and the arithmetic is not close. The honest caveat is that the cost of the software is rarely what decides these projects. The cost of connecting it to your booking system is.
What do people get wrong about full-duplex voice AI?
Three misconceptions cause most of the disappointment.
Misconception one: full-duplex means the AI is smarter.
It does not. Full-duplex is about the timing of the conversation, not the quality of the answers. A full-duplex agent connected to a bad knowledge base will interrupt you politely and then tell you something wrong.
Misconception two: it will handle Cantonese, English and Mandarin equally well.
Hong Kong calls switch languages mid-sentence, and mixed Cantonese-English is still the hardest case for every provider. Test with real recordings of your own customers before committing, not with a scripted demo.
Misconception three: you can deploy it in an afternoon.
The voice model is the easy part. The work is in the connections: your booking system, your stock data, your escalation rule for when a caller gets angry, and your record of what the agent promised. A voice agent with no escalation path is a complaint generator.
Is full-duplex voice AI ready for your business yet?
For narrow, repetitive, high-volume calls, yes. For anything involving a judgement call, money, or an upset customer, it should hand over to a human, and you should design that handover before you launch.
A sensible sequence looks like this:
--- Count how many calls you actually miss in a week. If the number is under ten, this is not your problem to solve first.
--- Pick one call type that has a factual answer.
--- Write down the exact sentence the agent says when it does not know. That sentence is the most important line in the whole deployment.
--- Record and review the first two weeks of calls. Every failure is a rule you did not write yet.
If your customers reach you mainly by message rather than by phone, the same logic applies to text channels, and the cost structure there changed this year too. Our guide to what WhatsApp replies now cost Hong Kong businesses covers that side.
The technology finally sounds human. Whether it helps your business depends on something much older than AI: knowing exactly which job you are hiring it to do. We understand AI. UD stands with you.
Reviewed by the UD AI team, Hong Kong.
Ready to work out whether voice AI fits your business?
Working out which calls to automate, and which must stay human, is the part worth getting right. UD has spent 28 years helping Hong Kong businesses make exactly this kind of decision, and we will walk you through it step by step, from assessing your call volume to going live.
