I Tried ChatGPT’s Improved Voice Mode and It Sounds More Natural Than Ever.
The latest update, currently rolling out to ChatGPT and available to all users, is GPT-Live. OpenAI describes it as a “next generation” of voice models for interacting with an AI-powered chatbot, and you might find that it will make you spend more time chatting than typing.
Voice mode is nothing new for ChatGPT, but it previously offered a relatively simple wrapper around standard text input and output. It was marketed as a more natural way to interact with AI, but GPT-Live promises to take this flexibility to an even higher level.
For the first time, voice mode will be able to think in the background while continuing a conversation. It will also give you extra space to pause when needed and will signal that it’s still listening with phrases like “mmm” or “yes.”
The update is expected to be available on mobile devices and the web version now (or very soon). Free users will get access to GPT-Live-1 mini, while paid users will get access to the even more intelligent GPT-Live-1 model.
How GPT-Live Works
OpenAI’s ultimate goal is to make conversations with ChatGPT feel like conversations with a real person, and GPT-Live brings us closer to that goal. Initially, interacting with AI via voice required a separate model for translating speech to text, another for responding to queries, and yet another for translating text to speech.
ChatGPT’s previous voice mode combined all of this into a single AI model, but it was still turn-based: you spoke, the chatbot responded, then you spoke again. With GPT-Live, ChatGPT can speak and listen simultaneously. You can interrupt it as needed, and responses should be faster and more accurate.
The new voice mode is claimed to be smarter at recognizing the difference between a pause in mid-thought and the actual completion of a request. The model now recalculates several times per second whether to “speak, continue listening, pause, interrupt, or launch a tool.”
An additional benefit of the update is that even complex work and deep thoughts can be transmitted to ChatGPT servers in the background while the conversation continues. You can also instruct ChatGPT to pause or slow down; visual responses have also been improved, so you can, for example, see pop-up cards with location, weather forecast, and sports scores.
You can now also ask GPT-Live to translate something into a foreign language while you speak. This new feature allows you to hear a continuous translation into the other language, without pauses or interruptions. Improvements have also been made to the ability to ignore background noise (such as traffic or nearby conversations).
GPT-Live Testing
To enable voice mode in the mobile app, tap the sound wave icon to the right of the prompt window. The new mode looks very similar to the old one, but after the update, you’ll see the word “Live” at the top of the screen (at least for now; you can tap it to switch back to the old modes).
It’s immediately clear that the improved voice mode sounds more realistic and natural. ChatGPT will speak in a varied and expressive manner, adding helpful prompts like “let me check” when searching for something. It will also occasionally pause and drawl.
I spent a few minutes chatting with GPT-Live about upcoming movies, the latest football matches, and tech news, and received thoughtful and respectfully brief replies (voice mode remains a safe haven for those who don’t want to see a bunch of text messages in every reply).
There were a few moments when speech was interrupted and the conversation froze, but these occurred within about half an hour of conversation (presumably these errors will be corrected over time). Interruptions are also handled well: the AI pauses to acknowledge what was said and then continues its thought.
You can adjust the level of intelligence ChatGPT puts into the new voice mode: tap the slider icon (in the upper right corner), then tap “Intelligence.” There are three modes to choose from— Instant , Medium, and High —with varying tradeoffs between response speed and response detail and accuracy.