Gemini音频模型全面升级：实时语音代理进入生产级可用时代
Improved Gemini audio models for powerful voice experiences
谷歌升级Gemini 2.5 Flash原生音频模型，显著提升函数调用、指令理解和对话流畅度。实时语音翻译功能正在美国、墨西哥和印度的安卓版谷歌翻译应用中推送测试，企业应用场景包括Shopify Sidekick和美威抵押贷款的AI助理。
Google enhanced Gemini 2.5 Flash Native Audio for live voice agents. Key improvements: sharper function calling, robust instruction following, smoother conversations. Live speech translation now rolling out in Google Translate app beta on Android in US, Mexico, and India. Use cases: Shopify Sidekick (merchants), United Wholesale Mortgage Mia (loan generation - over 14,000 loans generated). Available via Gemini API in Google AI Studio. Voice AI is becoming production-ready for customer-facing applications.
对 AI 行业的影响
语音AI长期以来受困于’玩具级’体验，难以支撑真实商业场景。Gemini音频模型在函数调用精度上的突破，意味着AI可以在通话过程中实时查询信息、执行操作，而不只是简单对话。Shopify和房贷领域已出现规模化商用案例，标志语音AI正在跨越从演示到生产的关键门槛。
原文参考
来源：Google DeepMind Blog · 2025-12-12