A talking avatar is judged almost entirely on one number, how long the silence lasts after you stop speaking. Below roughly a second it reads as a conversation. Above it, everyone starts talking over each other.The loop has five stages, and they are not equally expensive. Capture and voice-activity detection cost about 40 ms. Streaming speech-to-text returns a stable transcript around 180 ms after the last phoneme. First inference token lands near 220 ms. Text-to-speech emits its first audio chunk about 190 ms later, and rendering the first lip-synced frame costs another 160 ms.The important detail is that these stages overlap. We do not wait for a complete transcript before prompting, and we do not wait for a complete sentence before synthesizing. Partial transcripts are sent speculatively, and the first clause of the model's response is handed to text-to-speech while the rest is still generating.What is worth optimizing is the commit decision. Ending an utterance too early produces a confident answer to half a question, which is worse than a slightly longer pause, so we bias toward waiting and mask the difference with a short acknowledgement when a tool call is involved.