GPT-Live от OpenAI: как построили голосовой ИИ без пауз между репликами

OpenAI published a detailed account on its blog of how the company's engineers built GPT-Live — a system for continuous voice interaction with AI — in six months. The product is based on a turnless speech model and a low-latency infrastructure, which together deliver a faster and more natural dialogue between humans and AI.

What sets GPT-Live apart from conventional voice assistants

Most voice AI systems, including early versions of Voice Mode in ChatGPT, operate on a turn-based model. The principle is simple: the AI listens to the user, detects a pause, recognizes it as the end of a turn — and only then generates a response. This creates noticeable delays and makes it impossible to interrupt the system mid-sentence, which is atypical of natural conversation.

GPT-Live is built differently. According to the OpenAI blog, the system uses a turnless architecture: the model processes incoming speech continuously, without rigid boundaries between turns. The user can clarify, change topics, or interrupt at any point during the interaction — the system responds without a noticeable wait.

Key characteristics of GPT-Live as described by OpenAI:

• Turnless speech model — audio processing without fixed separation into user turns and AI responses
• Low-latency architecture — every layer of the stack is optimized for minimum response time
• Real-time streaming — audio is processed as a continuous stream, without queue buildup
• Six months — the time from architectural concept to production launch

Why the old system couldn't simply be improved

A standard voice pipeline consists of three sequential components: speech recognition (ASR), a language model, and speech synthesis (TTS). Each stage waits for the result of the previous one. This scheme is predictable and reliable, but fundamentally unsuited for continuous interaction: it is not designed for simultaneous bidirectional real-time audio streaming.

According to the OpenAI blog, engineers decided to build the real-time system from scratch rather than adapt the existing infrastructure. Latency became the primary constraint in the design: every component — from the model to deployment — was built with it in mind.

"GPT-Live delivers continuous voice interaction with AI, using a turnless speech model and a low-latency architecture for faster, more natural conversations," reads the official

OpenAI blog.

The six months from concept to launch cover the full engineering cycle: architectural decisions, prototyping, latency optimization, and production deployment.

How GPT-Live is already changing conversations with AI

GPT-Live powers the Advanced Voice Mode in ChatGPT — it is this system that drives the app's voice mode. For end users, the difference is tangible: instead of a conversation with pauses and waiting, there is a smooth, responsive dialogue. Interrupting the AI mid-sentence, asking a follow-up question in the middle, changing the topic — all of this now works without noticeable breaks, just like an ordinary phone call.

The turnless approach and low-latency architecture are setting a new engineering standard for voice AI. Companies developing voice operators for call centers, in-car assistants, and voice interfaces for people with disabilities will benchmark against this level when designing the next generation of their systems.

What this means

In six months, OpenAI reimagined the fundamental principle of voice AI — from turn-based interaction to continuous dialogue. GPT-Live is not an incremental improvement, but a shift in architectural paradigm. Response speed and conversational smoothness are no longer a bonus — they are becoming a baseline standard that the entire voice AI market will now have to meet.