OpenAI has made the GPT-Live-1.1 speech model available to developers, allowing them to integrate it into their own applications and create voice-based intelligent agents capable of conducting real-time conversations.

This model is not new at all. Since October 15th, both ChatGPT Plus and Pro users have been utilizing its voice functionality – in other words, it has been in use within these products for over a month before being officially launched for sale.

OpenAI emphasizes two aspects of these improvements: firstly, the ability to more reliably identify information such as letters and numbers, which requires precise understanding; secondly, instructions are now followed more consistently.

The first point may seem insignificant at first glance, but anyone who has worked in voice customer service knows how crucial it is.

Order numbers, mobile phone numbers, license plate numbers, and shipping label numbers are the areas where voice systems are most likely to make mistakes. A single error in distinguishing between “B” and “D”, or “15” and “50” can render an entire phone call futile, forcing customers to transfer the call to a human operator. Many voice customer service projects fail not because the algorithms are unable to engage in conversations, but because they simply cannot accurately remember such sequences of numbers.

Therefore, the most direct beneficiaries of this new approach are scenarios involving customer service, appointment scheduling, and after-sales support. The previous solution, which relied on a combination of “voice recognition, text modeling, and voice synthesis,” suffered from high latency and the possibility of errors in each individual component. An end-to-end model, on the other hand, can significantly reduce the overall processing time.

指令 compliance is stronger in this case, but this actually leads to another type of persistent problem: when you tell it to only respond to inquiries about returns and exchanges, it might start recommending products during the conversation.

This announcement does not provide information regarding pricing or latency figures. Before implementing the connection, it is recommended to conduct tests using actual recordings from your own business operations—especially those that contain accents or background noise, as such examples are usually not included in official demonstrations.