OpenAI’s bidirectional audio model shipped, and its audio already carries an inaudible signal. GPT-Live solves the voice. With Radius you build the data layer on the audio.
The Voice Is the Easy Part in a Bidirectional Audio Model
In March — when this article first ran — OpenAI was reportedly building a bidirectional audio model, codename BiDi, that could do something no voice assistant could: keep talking when you interrupt. On July 8, it shipped as GPT-Live: a full-duplex voice model that listens while it speaks, acknowledges you mid-sentence with “mhmm,” stays quiet while you think, and hands deeper work to a frontier model in the background without breaking the flow of conversation.
GPT-Live is now the default voice experience across ChatGPT’s tiers, and more than 150 million people already talk to ChatGPT every week. The voice-first device isn’t a rumor anymore either: OpenAI’s first hardware product with Jony Ive is reported to be a screenless smart speaker, targeted for 2027. Voice — not text, nor images or touch — is the interface OpenAI is building the company around. And voice runs on audio: on the speakers and microphones this bet is about to put everywhere.
GPT-Live makes voice AI feel human. Which means the thing almost nobody was talking about in March is now the thing that matters.
Voice Is the Interface. Audio Is the Infrastructure.
Tony Fadell caught the distinction on Lenny’s Podcast, describing the sensors an AI-era home actually needs to “pick up a lot of comings and goings and who’s who in the room and your voice and these kinds of things — audio, I should say, not voice.”
If voice is the primary interface, every device ships with always-on, well-engineered microphones and speakers: phones, cars, kiosks, wearables, rooms. Voice is why that audio infrastructure gets deployed. Audio is everything else it can carry.
Audio sensors is what an AI-era home actually needs.
Because a microphone listening for your voice is also a receiver. A speaker that answers you is also a transmitter. Once people routinely talk to devices in physical spaces, the audio channel between person and device is open — continuously, bidirectionally — and that channel can carry more than conversation.
An inaudible ultrasonic signal can travel on the same audio layer as the voice — encrypted, device-to-device, carrying data. Not as a second step: the medium that carries the conversation simultaneously carries the handshake — identity verification, proof of presence, transaction authorization. Three channels in a single ultrasonic pulse the human never hears.
With LISNR’s Radius, this data-over-sound transmission exists today. Until now, it has been solving problems in a world where the primary interface was still a screen. GPT-Live changes the context. When the conversation moves to voice, the data moves to audio.
GPT-Live Audio Already Carries an Inaudible Signal
Since July 31, audio generated with GPT-Live through ChatGPT Voice and the OpenAI API includes a SynthID watermark. SynthID is a provenance technology developed by Google DeepMind and adopted across the industry; OpenAI’s public verification tool detects its signal in supported audio files.
The watermark is not metadata attached to a file. It is encoded into the audio itself: the waveform is converted into a spectrogram, the pattern is embedded, the waveform is rebuilt. While human ears can’t detect it, machines can after transcoding, MP3 compression, speed and pitch shifts, and reportedly even after the audio has been played through a speaker and re-recorded by a microphone. That durability is its purpose. In OpenAI’s words, SynthID preserves a signal where metadata does not survive.
On August 2, Article 50 of the EU AI Act took effect, requiring AI-generated content to be marked in machine-readable form. An inaudible signal on the audio channel is now shipping practice for a voice AI with 150 million weekly users — and, in the EU, a legal requirement.
Bidirectional audio can carry more than conversation. It can carry data for many uses cases.
A watermark answers one question: did this audio come from this model? One fact, stamped at generation, verified afterward, and independent of anything said in the conversation. A handshake answers three, live: who is present, where they are, and what they are authorized to do. That takes data moving in both directions — device-to-device, encrypted, in real time, offline. The watermark identifies audio. The handshake transacts through it.
Retail and GPT-Live with Bidirectional Audio
A customer at a smart kiosk says: “I’d like to return this and exchange it for a different size.” GPT-Live handles the conversation beautifully — adjusting when the customer changes their mind mid-sentence. But today, the system still steps out of the voice interaction to verify identity: scan a QR code, search for an account, type in a loyalty number. The conversation happens on voice. The verification doesn’t happen on audio — yet.
Now that the channel is open, the verification can ride on it. An encrypted ultrasonic signal between the customer’s phone and the kiosk confirms identity and pulls up the purchase the moment the conversation begins. No code to scan, no account to search on a smartphone or the touchscreen of a kiosk. The handshake travels on the same medium as the voice, and the customer never leaves the interaction — authentication, payment credentials, and loyalty data in a single ultrasonic handshake.
Mobility and GPT-Live with Bidirectional Audio
A voice assistant in your car reroutes you, adjusts your schedule, pays a toll — all hands-free. GPT-Live makes that conversation seamless. But the toll still needs to know your car is at this terminal, right now, with certainty. GPS gets you close. Underground, in a parking garage, GPS gets you nothing. Today that means a transponder, a tag, or logging into a payment app afterward.
The audio channel is already open. If the same layer that powers the voice interaction carries an ultrasonic proximity signal — an automatic entry-exit handshake that verifies your vehicle at the terminal and authorizes payment — the extra device disappears. Sound penetrates metal and glass where Bluetooth and NFC can’t reach. The assistant confirms the charge, the ultrasonic handshake verifies the location, and you drive through. No transponder, no app — one medium doing both jobs.
Payments and GPT-Live with Bidirectional Audio
Your AI assistant walks you through a high-value purchase — conversationally, in real time, adjusting as you ask questions. GPT-Live makes that interaction feel natural. But how does the system verify that you — not someone who intercepted the voice stream — are physically present at the point of sale? Today you have a PIN, a biometric scan, a tap of a physical card. The voice does the conversation. Something else does the proof.
Now the proof can travel with the voice. An encrypted ultrasonic tone confirms the person authorizing the payment is physically at the terminal. One medium — the proof rides with the conversation. And because it works offline — no cloud lookup, no cell signal required — it works in the subway station, at the concrete stadium concession stand or turnstile, and in the underground garage where cloud-dependent systems fail.
In each case, GPT-Live creates the condition: a continuous, bidirectional audio channel between a person and a device. Data-over-sound creates the opportunity: the same channel carries the handshake.
Why the Conversation Company Won’t Own the Handshake
OpenAI already tried to own a transaction layer. Last fall they launched Instant Checkout, a feature that promised users could buy things “without ever leaving the chat.” It seemed like the logical extension: if the AI handles the conversation, why not let it handle the transaction too?
It didn’t work — it became known as “OpenAI’s shopping glitch” — and by March, OpenAI had quietly reversed course, routing shoppers back to merchant apps instead. The commerce layer — identity verification, payment processing, fraud prevention at the point of interaction — turned out to be a different problem than making conversation feel natural. Nothing in the months since has changed that. Building a great voice interface doesn’t mean you can also build the infrastructure that verifies who someone is, where they are, and whether the transaction is legitimate.
This isn’t a knock on OpenAI. It’s how they now design. GPT-Live doesn’t do its own deep reasoning — it delegates to a frontier model in the background. It doesn’t do its own provenance — it adopted a specialist’s watermark. The architecture of modern voice AI is decomposition into specialized layers, and OpenAI’s own launch is the proof. The physical-world verification stack — deterministic proof of presence, encrypted device-to-device authentication, offline transaction capability — is exactly such a layer. It belongs to the specialists whose entire focus is that handshake, not bolted onto a chat product.
Models will keep getting better at conversation. The handshake belongs to the data layer.
One Audio Channel: Conversation, Provenance, Verification
With OpenAI, voice is becoming the primary interface between humans and AI — and the conversation layer that makes everything else possible is shipping now. With Radius, you build the verification and interaction layer that rides on the same audio.
GPT-Live opened the channel. SynthID proved the channel carries an inaudible signal. What’s left is the layer physical-world applications have always needed: the handshake — actual data, not just a mark — on the medium the user is already engaged with. No extra hardware. No second step. No fumble.
NFC, QR codes, and Bluetooth beacons work — in interactions designed around screens and taps. In a voice-first interaction, where the user expects a seamless, hands-free experience, every switch to a different medium is friction. GPT-Live eliminates that friction for the conversation. Ultrasonic data-over-sound eliminates it for everything else, completing the stack where those technologies can’t reach: through metal, through glass, underground, offline, and across a room.
Ultrasonic data-over-sound is already deployed — verifying transit riders, enabling verified retail impressions, closing the loop between ad exposure and purchase. The range is software-defined, from a few centimeters for payment verification to thirty meters for venue-wide engagement, adjustable by API call, not by swapping hardware. Any device with a speaker or microphone becomes a secure transaction terminal within those deployments — and the voice-first era is installing those speakers and microphones everywhere, anyway.
The GPT-Live API is coming; the sign-up form is already open. The window before it opens is when the data layer gets designed — so that when voice-first products ship, the handshake is built in, not bolted on.
Voice is solved. Provenance is signed. Audio carries both. Now build the data layer on the same channel.
This article is part of LISNR’s ongoing exploration of how deterministic presence verification is reshaping commerce and media measurement. For the full Proof of Presence thought leadership series, start with Your Omnichannel Promise Has a Presence Problem (Article 1), then read “Trust Me, Bro” Is Not Proof of Presence (Article 2), and Mind the Proof Gap (Article 3).
Subscribe to LISNR’s LinkedIn newsletter “Buzz” so you don’t miss what’s next.