what services are you using in your pipeline? (stt, llm tts) and have you prompted the LLM to understand it's receiving transcribed text which may be inaccurate and to judge intent from context?