OpenAI rolls out more natural voice mode for ChatGPT

OpenAI has updated ChatGPT's voice feature with a new model called GPT-Live, marking a notable improvement in how natural spoken conversations with the assistant feel.

From three systems to one

ChatGPT's original 2023 voice mode relied on three separate systems: converting speech to text, generating a response, and converting that response back into speech. The later Advanced Voice Mode merged these into a single multimodal model. GPT-Live goes a step further with what OpenAI describes as a "full-duplex" architecture, allowing the model to listen and speak at the same time.

This addresses a common frustration with earlier versions, which would sometimes mistake a brief pause for the end of a sentence and begin answering prematurely. During conversations, ChatGPT may now vocalize short acknowledgements like "yeah" or "mhmm," similar to how a person responds while listening. GPT-Live can also hand off more demanding tasks, such as those requiring reasoning or research, to more capable models running in the background.

Access and customization

The GPT-Live model is available through the ChatGPT app on Android and iOS, as well as via any web browser. Paid subscribers on the Pro, Plus and Go plans get access to the full GPT-Live-1 model, while free-tier users receive the lighter GPT-Live-1 mini version. To start, users tap the waveform icon in the text box, granting microphone access on first use.

Within settings, users can choose between three intelligence levels — Instant, Medium and High — though this option is limited to paid plans. Voice conversations continue running even if the app is minimized or the device is locked. One current limitation is the lack of screen or video sharing support, though users can switch back to the older voice model through the settings menu if needed.