Part 1 of 4 · 1 min read
Overview
Voice removes the screen: no progress bar, no scrollback, no undo button. Timing, interruption and confirmation carry the whole experience, so latency budgets and turn-taking become correctness issues. It is a Specialist module: valuable for interviews at voice-focused companies and for understanding real-time constraints that also appear in streaming UIs (Level 11 module 4).
About the example, and what is real
Real: the turn-gap simulator, endpointing cut-in model, barge-in playback and history-truncation logic, filler policy, speech-time and voice-formatting helpers, and the 19 tests in code/mini-voice; the numbers come from running them. Assumed: every stage latency (600 ms endpointing, 150 ms finalisation, 450 ms model start, 60 tokens per second, 220 ms audio chunk, 80 ms network), the pause distribution, noise rate (3 bursts per minute), tool times, and a speaking rate of 165 words per minute. Not real: no audio, no speech models, no telephony, no human callers.