OpenAI Unveils GPT‑Live: ChatGPT Finally Talks Like a Real Person

Previous voice AI approaches

Early voice systems used a cascade architecture: speech‑to‑text, a large language model, then text‑to‑speech. This introduced latency, added complexity, and risked information loss between components. Subsequent round‑trip models such as Advanced Voice Mode processed audio within a single model, reducing delay but still required the user to stop speaking before the model could respond. The turn‑taking mechanism relied on silence detection, causing mis‑detections of pauses or background noise and leading to awkward interruptions.

GPT‑Live architecture

GPT‑Live adopts a full‑duplex design that processes incoming audio and generates outgoing speech concurrently. The model can emit short acknowledgments (e.g., “mhmm”, “yeah”), pause listening, or resume speaking at any moment, enabling continuous back‑and‑forth interaction and real‑time translation.

Continuous interaction is decoupled from deeper task processing. When a request requires web search, advanced reasoning, or stronger agent capabilities, GPT‑Live delegates the request to a back‑end model (currently GPT‑5.5). The delegation occurs while the front‑end maintains the ongoing conversation, allowing the system to handle multiple tasks without breaking the dialogue flow.

Evaluation

OpenAI built a human‑evaluation framework measuring pleasantness and fluency in 5‑ to 10‑minute one‑on‑one dialogs. Across dimensions such as overall preference, turn‑taking smoothness, interruption frequency, and naturalness, GPT‑Live‑1 and GPT‑Live‑1 mini were preferred over Advanced Voice Mode.

GPQA: GPT‑Live‑1 outperformed Advanced Voice Mode on expert‑level scientific reasoning (biology, chemistry, physics).

BrowseComp: GPT‑Live‑1 showed clear gains in web‑search and information‑finding tasks.

τ³‑Voice Telecom (internal): GPT‑Live‑1 surpassed Advanced Voice Mode in multi‑turn telecom customer‑service scenarios.

ChatGPT Voice experience

More than 150 million weekly users interact with ChatGPT via voice. With GPT‑Live, users can interject, pause, or ask the model to slow down; the model acknowledges listening with cues like “mhmm” or “got it”. Nine new voice skins have been added.

Answers can be generated with three reasoning‑strength options—Instant, Medium, High—allowing trade‑offs between speed and depth. Visual cards appear for weather, stocks, sports, and similar topics. The system better tolerates background noise and waits when the user needs time to think.

Rollout

GPT‑Live is being rolled out globally to ChatGPT users on iOS, Android, and chat.openai.com. GPT‑Live‑1 becomes the default model for Go, Plus, and Pro tiers; GPT‑Live‑1 mini is the default for Free users.

Reference: https://openai.com/index/introducing-gpt-live/