📄 中文摘要
在构建需要评估学生英语发音的 AI 代理时，直接将音频发送给大型语言模型（LLM）进行评分的做法并不可行。这是因为 LLM 只处理文本标记，而无法直接接收音频信号。当要求 LLM 从文本转录中评估发音时，实际上是在请求其从已经丢失所有声学信息的文本表示中推断声学特性。这导致 LLM 生成的分析虽然看似可信，但实际上是完全虚构的。因此，使用专门的语音 API 更为合适，可以直接处理音频信号并提供准确的发音评估。
为什么你的 AI 代理应该使用语音 API 而不是 LLM 推理
出处: Why Your AI Agent Should Use a Speech API Instead of LLM Inference
发布: 2026年2月22日
📄 English Summary
Why Your AI Agent Should Use a Speech API Instead of LLM Inference
Building an AI agent to evaluate a student's English pronunciation may tempt developers to send audio to a large language model (LLM) for scoring. However, this approach is flawed because LLMs process text tokens and do not directly handle audio signals. When asked to evaluate pronunciation from a transcript, the LLM attempts to infer acoustic properties from a textual representation that has already discarded all acoustic information. The outcome is a confident yet entirely fabricated analysis. Utilizing a specialized speech API is more appropriate, as it can directly process audio signals and provide accurate pronunciation assessments.