A voice agent's 2,500ms problem was scheduling, and one timer was 74% of the fix

Start with the arithmetic the post does not print. Moving from 2,500ms to under 1,800ms is a cut of at least 700ms, roughly 28 percent ^[14]. Replacing an 800ms silence timer with a neural voice activity detector that declares end-of-turn after 280ms accounts for 520ms of that on its own ^[6][16] - about three quarters of the whole improvement ^[16]. Nothing in that decision is a model decision. It is a decision about when you are permitted to believe the caller has finished.

The tuning is where the risk lives. A 280ms minimum silence window is fourteen frames at 20ms apiece, and the three frames of prefix padding retain 60ms of audio ahead of speech onset ^[17]. The detector opens at 0.65 and closes at 0.35, a hysteresis band 0.30 wide ^[7][21], which is what stops the state machine flapping inside a word. Shah says the shorter window did not produce false cutoffs when callers paused mid-thought ^[6]. That is a single-sourced claim with no error rate attached, and it is the one number a buyer would most want.

The savings also do not add up, and that is informative rather than damning. The two published wins are 520ms on turn detection and 600ms on time-to-first-audio ^[6][9], which together come to 1,120ms against a headline delta of 700ms ^[18]. Overlapping stages behave that way: a saving on one stage disappears behind another that is still running, and the author separately lists jitter on cellular networks, variance in LLM response time and buffering in TTS output as real-condition overhead ^[12]. The per-stage budget table the post refers to is not present in the text we were given ^[13], so the route to under 1,800ms cannot be reconciled line by line.

Then there is the path that undoes everything. A standard tool-calling roundtrip to check technician availability or look up an address routinely added 1,100ms ^[10]. Bolted onto a 1,800ms turn, that is 2,900ms, which is 400ms worse than the prototype the rewrite replaced ^[19]. Every hang-up the architecture fixed comes back the moment the agent has to ask the CRM anything. Shah says he implemented speculative pre-fetching, but the account we have breaks off mid-sentence ^[11], so what the system does when the speculation is wrong is not on record.

The barge-in handler is the part that should reshape a vendor shortlist. When speech is detected while the agent is speaking, the code clears the audio output buffer and aborts the in-flight generation before returning to listening ^[8]. That requires a text-to-speech stream you can flush and a model you can cancel. A faster model you cannot interrupt is worse on a phone call than a slower one you can, because the caller talking over stale audio is the failure everyone remembers.

And the ceiling stays where it was. Human turn gaps run 200 to 300ms, with 600ms already feeling awkward ^[3]. Under 1,800ms is three times the awkwardness threshold and six to nine times a natural gap ^[20]. The rewrite bought a caller who stays on the line. It did not buy conversation.