This is usually solved by splitting the output keep the markdown version for the UI and transcript and pass a cleaned version (without markdown symbols to the voice/TTS layer.