Say it wrong and keep going. An automated scorer counts hesitation, repetition and self-correction as costs in their own right, so stopping to repair a word usually loses more than the wrong word would have, and it is the one habit that good learners have spent years building.
That single instruction is worth more preparation than any amount of vocabulary here, because it runs against everything careful speakers do naturally. A person marking you might hear a repair as diligence. Nothing in this test can.
Eight facts about the format
- it is marked entirely by machine, which is the defining feature and the source of both its speed and its criticism.
- results arrive in days, sometimes within forty eight hours, which is far faster than the alternatives and is often the reason candidates choose it.
- scored out of 90, with separate scores by skill and by enabling skill such as pronunciation and grammar.
- it is a single two hour sitting with all four skills, rather than a speaking test on a separate day.
- the speaking is recorded and scored by algorithm, so clarity and pace matter more than personality.
- fluency is scored mechanically, which means hesitation and self correction cost marks even when the language is correct. Speaking steadily beats speaking accurately here.
- integrated tasks are heavily weighted, and one response can contribute to several skill scores at once.
- it is accepted for many visa routes as well as by universities, though acceptance varies by country and should be checked directly.
One response, several scores
The structural feature that surprises candidates is that responses are not scored against one skill each. Reading a prompt aloud contributes to both the speaking and the reading numbers. Writing a summary of something you listened to contributes to both writing and listening. A single weakness therefore shows up twice, and a strength does too.
The practical consequence is that the score report is harder to read than it looks. A low reading number may be reporting your delivery on a read-aloud task rather than your comprehension of anything, and a candidate who responds by practising reading comprehension will not move it. Look at which tasks feed a number before deciding what the number was telling you.
What an engine can and cannot count
It is strong on the things that are measurable in a recording: pace, pausing, the length and placement of silences, the range and accuracy of what was said, and how closely a written response matches what the source contained.
It cannot tell whether an answer was worth making. A fluent, well-paced response that slides gently off the topic scores better than it deserves; a thoughtful response delivered in four hesitant fragments scores worse. That is not a fault to complain about. It is a specification, and it tells you exactly what to rehearse: continuous delivery first, content second, in the opposite order from the one a person would want.
Where it stops: repairing in flight
The stall is self-correction, and it is a stall specifically for the strongest candidates. Somebody with careful English notices a wrong article, goes back, fixes it, and has now produced a repetition and two extra pauses in exchange for one corrected word. The exchange is a loss every time.
Nothing about it feels wrong from the inside, because in every other setting this speaker has ever been in, correcting yourself was the right thing to do. Teachers rewarded it, colleagues read it as care, and an engine reads it as disfluency.
The drill is uncomfortable and short: speak for two minutes and do not go back for anything, whatever you hear yourself say. Two weeks of that changes the score more than two months of vocabulary.
Finding out whether a machine can follow you
Google Docs voice typing is free in a browser and it transcribes continuous speech rather than single words, which makes it the closest free approximation of being listened to by software. Read a paragraph into it, then read the transcript, and the places where it broke are the places your delivery stopped being machine-readable.
It replaced guessing at whether a machine can understand you, which is not something a person is able to tell you: a human listener repairs your speech automatically as they hear it and has no access to what they repaired.
Its limitation is worth stating precisely. It is a different engine from the one scoring the exam, and it is tuned to be forgiving, so a clean transcript is weak evidence and a broken one is strong evidence. Treat every failure as real and no success as a result.
What it looks like when it goes wrong
The recognisable failure is a candidate with excellent English and a mediocre speaking number, which almost always turns out to be pace and repair rather than anything about the language.
The second is a response that stops early because the candidate finished what they had to say. Silence at the end of a timed slot is measured, so filling the time with something relevant is worth more than a concise answer that leaves eight seconds of nothing.
The reason a human-marked test rewards a different performance entirely, and why preparing for one while sitting the other loses band, is what each test scores, and how to prepare for yours, and the human-marked alternative most candidates are choosing between this and is the four papers, the band scores, and what examiners reward.
What continuous delivery actually depends on, once the exam is out of the way, is not a technique for a test. It is how much of what you say is retrieved whole rather than assembled, which is the same mechanism under every fluent speaker: speed without accuracy, accuracy without speed, and how they meet.