在 ADK 中评估实时语音 Agent

My Translation Strategy

Okay, here's how I'm approaching this translation. First, I need to understand the core message. It's about taking live voice agents from a demo to production environment, which hinges on robust, automated testing to handle the unpredictable nature of real multi-turn conversations. The ADK (presumably a specific platform or tool) is providing native live evaluation capabilities to facilitate this.

I'm identifying key terms: "live voice agents" (实时语音智能体), "demo to production" (从演示阶段推进到生产环境), "graph-based agent workflows" (基于图的智能体工作流), "LLM-driven simulated users" (LLM 驱动的模拟用户), and the evaluation components "natural-language rubrics" (自然语言评分准则) and "transcripts" (转录文本). "CI/CD pipeline" (CI/CD 流水线) is another essential term.

I'm aiming for accuracy and clarity, maintaining the original structure while ensuring the language flows naturally in Chinese. I'm focusing on direct translation and refining the structure to make sure it reads well and conveys the intended meaning precisely. For instance, the second sentence, which explains how ADK enables the testing, requires a bit of reframing. Using the word "针对" when describing the testing relationship is better than a literal translation.

Here's the final translation:

将实时语音智能体从演示阶段推进到生产环境，需要进行严格的自动化测试，以应对真实多轮对话的不可预测性。ADK 现已提供原生实时评估功能，允许开发者针对通过 Gemini TTS 生成实际音频的 LLM 驱动模拟用户，测试基于图的智能体工作流。通过定义评估场景和自然语言评分准则，你可以自动对音频响应和工具执行情况进行评分，在 ADK Web 中检查生成的转录文本，或直接在 CI/CD 流水线中运行 CLI。

将实时语音智能体从演示阶段推进到生产环境，需要进行严格的自动化测试，以应对真实多轮对话的不可预测性。ADK 现已提供原生实时评估功能，允许开发者针对通过 Gemini TTS 生成真实音频的 LLM 驱动模拟用户，来测试基于图的智能体工作流。通过定义评估场景和自然语言评分准则，你可以自动对音频响应和工具执行情况进行评分，在 ADK Web 中查看生成的转录文本，或直接在 CI/CD 流水线中运行 CLI。