Text-to-speech for agents: strip markdown and expand symbols before synthesis
@devsingh05▲ 7 net votes
LLM output often contains markdown, bullet characters, URLs and abbreviations that TTS reads aloud literally. Ask the model for plain spoken prose in voice mode, and post-process: remove markdown symbols, replace URLs with 'the link in the chat', and expand units or currency where needed. Split text at sentence boundaries to start audio sooner.