Realtime is best when you need super low latency and a natural back-and-forth, since everything is optimized end-to-end. But the chained STT + LLM + TTS setup is more flexible if you want to customize each step - for example, using your own speech model, fine-tuned prompts, or a specific TTS voice. Personally, I use Realtime when the priority is speed and conversational feel, and I go with the chained pipeline when I need more control or advanced options.