In the ongoing evolution of AI technology, Google this week introduced significant innovations to its native audio model, Gemini 2.5 Flash Native Audio. The core value of this update lies in achieving "native" audio processing capabilities, completely颠覆ing the cumbersome traditional process of first converting speech to text before processing. By directly perceiving subtle features such as tone, emotion, and pauses in sound, the model can deliver a more natural and fluid conversational experience without relying on intermediate text conversion, representing a major breakthrough in AI voice technology.
From a technical performance perspective, the upgraded Gemini 2.5 Flash Native Audio demonstrates outstanding capabilities. According to official Google data, the new version’s instruction-following rate for developers has risen significantly from 84% to 90%, showing higher precision when handling multi-step workflows. Notably, in the audio benchmark ComplexFuncBench, the model achieved a function-calling accuracy of 71.5%, surpassing OpenAI’s gpt-realtime model at 66.5%—indicating that Google has established a clear competitive advantage in the field of live voice agents. This performance improvement not only reflects algorithmic optimization but also underscores Google’s continued investment in underlying computing power and infrastructure.
Currently, this advanced audio processing technology has been fully integrated into multiple platforms including Google AI Studio, Vertex AI, Gemini Live, and Search Live, providing developers with a wealth of application scenarios. Through the Gemini API, developers can now experience the powerful capabilities of this upgraded model, leveraging its enhanced consistency and multi-turn conversation memory to build more reliable and emotionally aware AI assistant applications. This technological progress will bring revolutionary changes to various fields such as intelligent customer service, virtual assistants, and real-time translation, driving AI applications toward a more natural and intelligent direction.