In short
An AI mock interview is a practice job interview where an AI plays the interviewer, asks real interview questions out loud, and marks how you answered. NativeSpeech runs a fixed set of questions, follows up on what you actually said, and scores the transcript against a published rubric — a band from 1 to 4 on 6 dimensions, each one quoting the sentence of yours it was judged on.
It runs in all 62 languages the conversation feature supports, and the report comes back in the language you practised.
It is an exam, not a lesson. Nothing is corrected while you are in the room, no score appears until it is over, and every learner sitting the same simulation is asked the same questions in the same order — which is the only way two attempts, or two people, can be compared at all.
The mark scheme
Each dimension is banded 1 to 4 and carries a fixed weight. The weighted average is your composite out of 100. The model grades three of the six from your transcript and must quote it; the other three are counted from the recording in code, because a model asked to count fillers will confidently invent a number.
Band 4: All four parts stated, the learner's own contribution is unambiguous, and the outcome is named.
Band 4: Position first, reason second, every time. Nothing in the answer is there twice.
Band 4: Steady and listenable from the first answer to the last.
Band 4: Rare enough that a listener would not notice them.
Band 4: Accurate and appropriately formal from start to finish.
Band 4: Consistently clear — nothing in the session would need repeating.
The evidence
Every band judged from your transcript arrives next to a verbatim sentence you actually said. Before that quote is shown it is checked against the recording word for word; anything the model tidied up, paraphrased or invented is dropped rather than displayed, and the report tells you how many quotes survived.
This is the part that makes a machine score worth anything. A number with no working shown is an opinion, and the first time a learner is told they said something they did not, every other line of the report stops being believable.
Beyond interviews
There are 6 scored simulations across 4 tracks, and each track is marked on what that room actually needs rather than on a single generic scale.
Marked on STAR structure: whether the situation, your own task, the action you took and the result are all there. Answers that stop at what you did, with the outcome implied, cap at band 3.
Marked on yesterday, today and blockers — in that order, with the blocker stated as a blocker rather than mentioned when asked. The lead interrupts for specifics.
Marked on impact before cause, on numbers instead of adjectives, and on whether you committed to a time for the next update. Detail without impact is band 1.
Marked on acknowledging the frustration, taking ownership and landing a next step the customer accepts — in that order. Jumping to the fix first is what keeps a call hot.
How it works
Choose a job interview, a daily standup, an incident call or a customer escalation, in any of 62 languages. There are 6 scored simulations across 4 tracks, each with its own published mark scheme.
A named interviewer works through a fixed set of questions and follows up on what you actually said. Nothing is corrected while you are in the room, and no score is shown until it is over — the same conditions as the real thing.
You get a band from 1 to 4 on 6 dimensions and a composite out of 100. Every band judged by the model quotes a verbatim line from your own answers; pace, filler rate and pronunciation are counted from the recording rather than judged.
Retake the same simulation and the report shows the movement dimension by dimension — "answer structure: 2 to 3". A composite of 75 is the readiness bar: every dimension at band 3 on average.
What it costs
Your first simulation is free — the whole thing, graded report included, and no account needed to sit it. A learner who has read one report knows exactly what they would be buying, which is a better pitch than any description of it.
After that a simulation is billed on the same clock as conversation practice: one credit per 5 active minutes, covering the microphone, the interviewer's voice, the model turns and the graded write-up. There is no separate per-session fee — answering in three sentences should not cost the same as rambling for ten minutes.
Practising without a mark scheme first? AI conversation practice runs the same engine and corrects you as you go, rather than holding it all back for a report.
Questions
