OpenAI has announced the launch of "GPT-Live-1" for its API, a voice model designed for developers to build voice-enabled applications and business workflows.

GPT-Live-1 enables "full-duplex" conversations, where listening and speaking can occur simultaneously. Conventional voice agents have faced challenges with latency and loss of context because they chained speech recognition (ASR), inference, and speech synthesis as separate processes. GPT-Live-1 simplifies these layers by handling both listening and speaking within a single model.

Additionally, it features a "delegation" function that allows conversations to continue even if a user interrupts, without interrupting the inference or task execution occurring in the backend. Developers can freely select the inference models and tools used in the backend, separately from the voice model governing the conversation. For example, it is possible to use a low-cost model for high-volume booking management while combining it with a model possessing advanced inference capabilities for complex customer support. Furthermore, ASR transcriptions and response texts are provided natively.

The API pricing is set at $0.05 per minute for the voice layer portion. Fees for the models and tools used in the backend are charged separately.

Source:

Source: