📄 中文摘要
语音代理在进行自由形式对话时，其行为具有非确定性特征。相同的代理在相同的提示下，每次都会产生不同的对话，传统的基于断言的测试方法在没有单一正确输出的情况下失效。因此，需要一种能够理解意图的评估工具。Voicetest通过将大型语言模型作为评判者来解决这一问题。该工具模拟与语音代理的多轮对话，并将完整的对话记录传递给评判模型，根据成功标准进行评分。该方法的核心在于三模型架构的设计，确保评估的准确性和有效性。
使用大型语言模型评估语音代理：工作原理
出处: Voice Agent Evaluation with LLM Judges: How It Works
发布: 2026年2月19日
📄 English Summary
Voice Agent Evaluation with LLM Judges: How It Works
Voice agents exhibit non-deterministic behavior during free-form conversations, producing different dialogues even with the same prompt. Traditional assertion-based testing fails in scenarios where there is no single correct output, necessitating an evaluator that understands intent rather than mere string matching. Voicetest addresses this challenge by employing LLM-as-judge evaluation, simulating multi-turn conversations with the voice agent and passing the complete transcript to a judge model that scores it against predefined success criteria. The approach is built on a three-model architecture, ensuring accurate and effective evaluation.