MockLive is a simulation platform for the conversations that decide things — a sales call, a piece of difficult feedback, a negotiation, an interview, an exam. The products differ. The hard part underneath them is the same one: making a spoken exchange feel like a conversation rather than two people taking turns at a radio.
Almost all of that difficulty is in the seam between turns. What follows is what we actually do about it.
Most voice systems end your turn by counting silence. That forces a choice with no good answer: a short threshold interrupts anyone who stops to think, and a long one puts dead air after every sentence.
In our production logs, 56% of the moments where a speaker's audio stopped were mid-answer pauses, not finished turns. A timer tight enough to feel responsive would therefore be interrupting the speaker most of the times it fired.
We use a semantic turn model instead — one that decides from what was said rather than how long it has been quiet. It holds the floor through a three-second pause in the middle of a sentence and releases straight away on a finished thought. To a silence timer those two things are identical at every threshold; that is the whole reason the category feels robotic.
Writing a reply and synthesising its first sentence both take provider time, and the second cannot start until the first has produced something to say. Waiting until you have finished speaking to begin that work puts all of it between your last word and the first one you hear back.
So the reply is written and synthesised while you are still trailing off, on the turn model's early signal, and is usually sitting in memory by the time your turn genuinely ends.
A speculative reply is generated from a snapshot of the conversation. If you then say anything else, that reply is answering something you had not finished saying — which sounds like not being listened to, and is a worse failure than being slow. Every speculation carries a key derived from the exact state it came from and is discarded unless that state is unchanged. There is no close-enough branch, and there must never be one.
Being cut off, redirected and pushed is a large part of what a high-stakes conversation actually is, so a simulator that cannot be interrupted is not simulating one. The obvious implementation breaks immediately: on laptop speakers rather than headphones, echo cancellation does not hold, the microphone hears the simulated voice, and that bleed is transcribed as the user speaking. Two lines is enough to start a loop where it interrupts itself.
The system therefore keeps listening while it speaks but does not act on the first thing it hears — it distinguishes its own voice bleeding back from a real interruption before yielding the floor. Deafness was the safe option and it cost the product the single behaviour it exists to rehearse.
Mid-turn replies land in about 400 milliseconds when speculative prep hits — measured end-to-end from your last syllable to the first word you hear. That number is not a single stage in the pipeline — it is the gap from the end of your audio to the first syllable you hear, including every deliberate wait.
While you are still trailing off, the turn model's early signal triggers reply writing and lead-sentence synthesis. When your turn genuinely ends, the audio is usually already in memory.
The live model streams tokens. As soon as the first complete sentence arrives, text-to-speech starts — overlapping the two slowest provider calls instead of running them back to back.
On a high-confidence early end-of-turn, playback can start before final confirmation, with a brief holdback so a wrong guess is cancelled before you hear anything.
The first sentence plays while later sentences synthesize during speech, so you never wait for a full paragraph to render before hearing a word.
The same engine powers Job interviews, IELTS Speaking, Roleplay, Reflect. Timed exam sections (PTE, TOEFL, DET) share the voice stack for prompts and scoring but use published exam clocks rather than free-form turn-taking.
Sales discovery, objection handling, manager feedback, negotiation, de-escalation — a counterpart who does not simply agree, and a scored debrief.
IELTS, TOEFL, PTE and the Duolingo English Test, on the published timings, marked against the published rubrics.
Resume-aware questions, live coding, follow-ups that go where your answer went, and a report the moment you hang up.
Think a decision through out loud. Session-scoped, deleted by default, no cross-session memory.
Because it is counting silence rather than understanding speech. A fixed endpoint timer has to pick a threshold: short, and it cuts off anyone who pauses mid-thought; long, and every exchange has dead air at the end. In our production logs 56% of the points where a speaker's audio stopped were mid-answer pauses rather than finished turns, so a timer set tightly enough to feel responsive would interrupt the speaker most of the time it fired.
A turn model that decides whether a speaker has finished based on what they said, not on how long they have been quiet. It holds the floor through a three-second mid-sentence pause and releases immediately on a completed thought — a distinction a silence timer cannot make at any threshold, because both look identical to it.
Writing and synthesising the reply while the speaker is still trailing off, so it is already in memory when their turn genuinely ends. The risk is answering a question that was not finished. Every speculation is therefore tagged with a key derived from the exact conversation state it was generated from, and it is only ever used if that state is unchanged — there is no close-enough branch.
The fastest way to judge any of this is to pause mid-sentence and see whether it waits. The first session is free.