The latest upgrade being pushed out to ChatGPT, heading to all users now, is GPT‑Live. OpenAI is describing it as a “new generation” of voice models for interacting with the AI chatbot, and you might find that it leads you to spend more time chatting than typing.
Voice mode for ChatGPT is nothing new, but previously it’s been a relatively basic wrapper on top of the standard text input and output. It has been billed as a more natural way to engage with the AI, but GPT-Live promises to dial this fluidity up to an even higher level.
For the first time, the voice mode will be able to think in the background while continuing the conversation. It’ll also give you extra space to pause when you need it, and indicate it’s still listening with phrases like “mhmm” or “yeah.”
You should find the upgrade on mobile and the web now (or very soon). Free users get access to GPT‑Live‑1 mini, while those on paid plans are able to access the even smarter GPT‑Live‑1 model.
How GPT-Live works
OpenAI’s end goal is to make talking to ChatGPT feel like talking to a real person, and GPT-Live gets closer to that. Originally, interacting with the AI via voice required a specific model for speech-to-text, another for actually responding to the query, and another for text-to-speech.
The previous voice mode in ChatGPT combined all of that into a single AI model, but it was still turn-based: You spoke, the chatbot answered, then you spoke again. With GPT-Live, ChatGPT can be talking and listening at the same time. You can interrupt it as and when needed, and responses should be faster and more nuanced.
The new voice mode is supposedly smarter when it comes to recognizing the difference between you pausing mid-thought and actually finishing your query. The model now recalculates several times a second “whether to speak, continue listening, pause, interrupt, or invoke a tool.”
Credit: OpenAI
An added benefit of the upgrade is that even complex work and deep thinking can be passed back to ChatGPT’s servers in the background, while the conversation is continuing. You can also tell ChatGPT to take a beat or slow down; visual responses have been improved as well, so you might, for example, see pop-up cards for locations, weather forecasts, and sports scores.
You can also now ask GPT-Live to translate something into a foreign language as you speak. Thanks to the new capabilities, you’ll hear a running translation in the other language as you talk, with no pauses or interruptions. Improvements have also been made in terms of ignoring background noise (like background traffic or conversations happening nearby).
What do you think so far?
Testing out GPT-Live
To get to voice mode in the mobile app, tap the soundwave-style icon to the right of the prompt box. The new mode looks a lot like the old one on the surface, but with this update, you should see Live at the top of the screen (for the time being, at least, you can tap this to switch back to the older models).
Right away, the upgraded voice mode feels more realistic and natural. ChatGPT will talk in a varied and expressive way, throwing in useful markers like “let me check” whenever it’s looking something up. It’ll lso hesitate and draw words out at times.
I chatted with GPT-Live for several minutes about upcoming movies, recent soccer matches, and tech news headlines, and got back answers that made sense and were respectfully brief (voice mode continues to be a refuge for those who don’t want to see walls of text for every response).
There are three ‘intelligence’ levels to choose between.
Credit: OpenAI
There were a couple of moments where the speech glitched and the conversation hung, but that was in about half an hour of chatting (presumably these bugs will get ironed out over time). Interruptions are handled well too, with the AI pausing to acknowledge what you’ve said and then continuing its train of thought.
You can tweak the level of thinking ChatGPT puts into the new voice mode: Tap the sliders icon (top right), then tap Intelligence. There are three modes to pick from—Instant, Medium, and High—with varying levels of trade-off between the speed of the response and how detailed and accurate it is.
