E2E Avg — the whole round trip: dealer stops speaking → agent's voice starts. This is the silence they actually hear. Under ~2s feels conversational; past ~4s they start talking over the agent or assume the line dropped.
E2E P95 — the slow tail. 95 out of 100 replies were faster than this. Judge the agent by this, not the average: one 9-second gap is what makes a dealer hang up, and averages hide those.
STT Avg — Deepgram turning their speech into text (Nova-3 batch or Flux streaming, per your env).
LLM Avg — the hosted model (Groq Llama 3.3 70B by default) composing the reply, including network round-trip and any provider rate-limiting.
TTS Avg — Deepgram Aura synthesizing the reply audio.
The three legs will not add up to E2E, and that gap is expected, not a bug. Between the dealer falling silent and STT starting, the agent waits to confirm they are really finished (turn endpointing, currently 0.4–1.5s — see turn_handling in src/sdr_agent/session.py). Network transit adds more. So E2E − (STT + LLM + TTS) is roughly your endpointing overhead: if that remainder is large, tune the endpointing delays; if a single leg dominates, that is the one to optimise.