OpenAI 最近发布的 GPT-Live 语音模型透过在同时说话和聆听时顺畅地处理诸如「um」和「er」之类的口语特质，在人机互动方面引入了重大进展。不同于以往在处理打断和背景噪音时会遇到困难的僵化、轮流对话的代理，这款新 AI 完美地处理了失误。透过持续分析音讯输入而不是将它们视为独立的讯息区块，GPT-Live 允许使用者在 AI 说话时插话而不会使对话偏离轨道，这与遇到杂音时会突然停止说话并错误地恢复为聆听状态的旧系统形成了鲜明对比。
GPT-Live 的一个显著特征是它对话语标记的使用，在对话期间于背景安静地发出诸如「cool」或「got it」的肯定语。虽然这可能对某些人来说很烦人，但这些提示在功能上模仿了人类的积极聆听，并作为 AI 保持其连接的实用指标，这在汽车等具有多变蜂窝接收讯号的环境中特别有用。这项能力代表了向前迈出的关键一步，因为 OpenAI 报告指出每周有超过 150 million 人与 ChatGPT 交谈，这突显了业界的信念，即语音将成为解锁新软体和设备生态系统的主要介面。
从历史上看，将自然对话数位化的尝试，从 Apple 1987 年的愿景影片到 2010 年代如 Siri 和 Alexa 的语音代理，在很大程度上忽略了人类语音混乱的现实，迫使使用者以完美的句子进行交流。虽然像 Google 的 Gemini 这样的现代迭代非常聪明，但它们仍然在杂音上遇到困难，并且需要手动设定以决定是否允许中断。GPT-Live 透过拥抱人类的弱点和非结构化对话来弥合这一差距，最终摆脱了过去生硬、有序的交流，转向流畅且反应灵敏的对话体验。
OpenAI's recently released GPT-Live voice models introduce a significant advancement in human-computer interaction by smoothly processing verbal idiosyncrasies like "um" and "er" while speaking and listening simultaneously. Unlike previous rigid, turn-by-turn conversational agents that struggled with interruptions and background noise, this new AI handles miscues flawlessly. By continuously parsing audio inputs instead of treating them as separate message chunks, GPT-Live allows users to talk over the AI without derailing the conversation, marking a stark contrast to older systems that would abruptly cease talking and erroneously revert to listening when encountering stray sounds.
A notable feature of GPT-Live is its use of discourse markers, quietly uttering affirmations such as "cool" or "got it" in the background during conversations. While potentially annoying to some, these cues functionally mimic human active listening and serve as practical indicators that the AI maintains its connection, particularly useful in environments with variable cellular reception like cars. This capability represents a crucial step forward, as OpenAI reports that more than 150 million people talk with ChatGPT each week, underscoring the industry's belief that voice will become a primary interface unlocking new software and device ecosystems.
Historically, attempts to digitize natural conversation, from Apple's 1987 vision video to the voice agents of the 2010s like Siri and Alexa, have largely ignored the messy reality of human speech, forcing users to communicate in perfect sentences. While modern iterations like Google's Gemini are incredibly smart, they still struggle with errant sounds and require manual settings to determine if interruptions are allowed. GPT-Live bridges this gap by embracing human foibles and unstructured dialogue, finally moving away from the stilted, orderly exchanges of the past toward a fluid and responsive conversational experience.