OpenAI Unveils GPT‑Live: Toward Human‑like Real‑time Voice Conversations

Previous voice‑AI architectures

Early voice systems used a cascade architecture : a speech‑to‑text model transcribed user audio, a large language model generated a response, and a text‑to‑speech model rendered the reply. This enabled voice interaction with cutting‑edge AI but introduced extra latency and potential information loss because each stage passed data to the next.

Subsequent round‑based models (e.g., Advanced Voice Mode) merged the pipeline into a single model, reducing delay and producing smoother turns. However, they still required the user to finish speaking before the model could respond. Turn‑taking relied on detecting silence, so brief pauses or background noise could be misinterpreted as the end of an utterance, causing premature interruptions.

GPT‑Live architecture

Full‑duplex continuous interaction : GPT‑Live processes incoming audio and generates outgoing audio simultaneously. At each moment the model decides whether to speak, keep listening, pause, interject, or invoke an external tool. This design removes the strict “one‑turn‑after‑the‑other” constraint and supports real‑time behaviors such as acknowledgments (“mhmm”, “yeah”) and live translation.

Delegation mechanism for deep tasks : When a request requires web search, complex reasoning, or advanced agent capabilities, GPT‑Live forwards the request to a backend model (initially GPT‑5.5). The front‑end maintains conversational flow while the backend performs heavyweight computation. Because the backend is a separate model, it can be upgraded independently as newer frontier models become available.

Human‑centric evaluation

OpenAI built a new evaluation suite that measures enjoyment and smoothness in 5‑ to 10‑minute one‑on‑one dialogues. The suite records five dimensions: overall preference, turn‑linking, interruption frequency, dialogue fluidity, and perceived naturalness.

In paired comparisons against Advanced Voice Mode, both GPT‑Live‑1 and GPT‑Live‑1 mini were preferred across all five dimensions.

Benchmark performance

GPQA – expert‑level scientific reasoning across biology, chemistry, and physics. GPT‑Live‑1 achieved a statistically significant lead over Advanced Voice Mode.

BrowseComp – agent‑style web‑search and hard‑to‑find information retrieval. GPT‑Live‑1 showed a clear improvement.

τ³‑Voice Telecom (internal) – multi‑turn telecom customer‑service tasks. GPT‑Live‑1 outperformed Advanced Voice Mode.

Deployment

GPT‑Live is being rolled out globally to ChatGPT users on iOS, Android, and chat.openai.com. Two variants are released:

GPT‑Live‑1 – default for Go, Plus, and Pro tiers.

GPT‑Live‑1 mini – default for Free tier users.

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