OpenAI has launched GPT-Live, a native voice model powering ChatGPT's conversational mode, announced with latency under 300 milliseconds and the ability to convey emotional nuance. Its technical hallmark: it removes the text intermediate step, which until now was the bottleneck.
Two weeks ago, we wrote that voice was becoming the primary interface, with seven billion raised in a single quarter by the sector. Here is the technical building block that makes this shift credible.
How it worked before
A classic voice assistant chains together three separate systems.
First, a transcription model converts your speech into text. Then, a language model reads that text and produces a written response. Finally, a speech synthesis model turns that response into sound.
Each step adds its own delay, and the whole pipeline typically produced a second or more before the first syllable. That's not much, but it's enough to stop a conversation from being one: you don't interrupt, you don't pick up the thread, you wait your turn as if talking to an answering machine.
The second problem is subtler and more interesting. Transcription throws away information. Your hesitation, your ironic tone, your irritation, the fact that you speak faster because you're in a hurry: none of that survives the conversion to text. The model receives bare words.
What a native voice model changes
A native system processes sound directly and produces sound directly. There is no intermediate conversion, so no cumulative delay, and crucially no information loss.
Under 300 milliseconds, you enter the ballpark of human conversation, where typical reaction time sits around 200 milliseconds. That's the threshold at which an exchange stops being turn-taking and becomes a dialogue: you can interrupt, react, hesitate together.
It also partially addresses a limitation we described in our article on humour: a lot of irony is carried by intonation, and a system that hears the tone has information that text simply doesn't contain.
A system that perceives the emotional nuances of your voice collects a different kind of information than your words. Fatigue, tension, emotional state are audible. We noted this in relation to what your AI knows about you: the structure of an interaction often says more than its content. With voice, that information becomes far richer. That's not a reason to avoid it; it's a reason to know what's being captured.
The trickiest point
A fluid, responsive, tone-sensitive conversation mechanically produces a much stronger impression of presence than a written exchange.
This is exactly the mechanism we described in our article on anthropomorphism: language is the most powerful signal of inner life we know, and voice amplifies it considerably. Knowing you're talking to software doesn't switch off that reflex.
It makes the questions we raised about loneliness and about attachment to AI companions more concrete. It's not the model's capability that changes the game; it's the fluidity of the exchange.
What to take away
This is real technical progress, and its most indisputable benefit is the one we highlighted in our article on accessibility: for someone with visual impairment, an older person, or someone unable to use a keyboard, a genuinely fluid voice interface isn't a convenience but a means of access.
Just keep in mind what's happening at the same time. A system that responds in 300 milliseconds, hears your tone, and adjusts its own is very pleasant to use and very easy to mistake for someone. The two things come together, and they don't come apart.