Google’s AI, Gemini , already makes talking to a virtual assistant feel more natural than it did when commanding Google Assistant. But nothing is perfect, so there’s always room for improvement. Gemini is now getting a few updates to improve its conversational quality.
Today, Google announced that it’s rolling out an update to Gemini 2.5 Flash Native Audio for live voice agents. For this update, Google focused on three key areas:
Sharper function calling: Reliability has been improved when triggering external functions. Gemini can now more accurately identify when to gather real-time information during a conversation. It can then seamlessly include that data in its audio response, without breaking the flow of the conversation.
Reliability has been improved when triggering external functions. Gemini can now more accurately identify when to gather real-time information during a conversation. It can then seamlessly include that data in its audio response, without breaking the flow of the conversation. Instruction following: The model now has a 90% adherence rate to developer instructions, improving over the previous 84%. As a result, it can better handle complex instructions and deliver more reliable outputs.
The model now has a 90% adherence rate to developer instructions, improving over the previous 84%. As a result, it can better handle complex instructions and deliver more reliable outputs. Smoother conversations: Gemini 2.5 Flash Native Audio can now retrieve context from previous points in the conversation more effectively, allowing for more cohesive conversations.
In addition to these improvements, Josh Woodward, VP of Google Labs, Gemini, and AI Studio, shared two other enhancements. One improvement makes it so that Gemini Live will not cut you off mid-sentence if you pause for too long. The other allows you to mute your microphone while Gemini Live is talking, so you don’t accidentally interrupt it.