Comment by artur44

A simple way is to split the model’s output stream before TTS. Reasoning/structured tokens go into one bucket, actual user-facing text into another. Only the second bucket is synthesized. Most thinking out loud issues come from feeding the whole stream directly into audio.