subtopic

STT → LLM → TTS (cascade) vs speech-to-speech

Learn the standard pipeline: the caller's audio is turned into text by speech-to-text (STT), the LLM decides what to say (and may call tools), text-to-speech (TTS) turns the reply into audio. Every block streams so the agent can start speaking quickly. Newer speech-to-speech (realtime) models do all three in one model: more natural and faster, but less control, harder to debug and often costlier. Most business agents still use the cascade.