A minimal pair drill is worth doing for the contrasts your first language does not make, and for nothing else. The rest of a generic pair set is a test you were always going to pass, and in most apps the rest is nearly all of it.
So the first job is not drilling. It is working out which four or five pairs are actually yours, which takes an afternoon and saves months. English holds a long list of sounds apart; your ear already holds most of that list apart; the intersection is small and it is the entire target.
Eight pairs, and how to use them
A minimal pair is two words differing in exactly one sound. Hearing the difference has to come before producing it, so listen first and only then repeat.
- ship and sheep, the short and long i, and the most useful pair in the language for most learners.
- live and leave, the same contrast in a pair that appears constantly in ordinary sentences.
- bat and bad, testing the final consonant, which is voiced in the second and lengthens the vowel before it.
- thin and sin, the th against the s, which is the substitution French and German speakers make.
- light and right, l against r, the pair Japanese and Korean speakers work hardest on.
- cheap and jeep, unvoiced against voiced, which is a difficulty for Arabic and Spanish speakers.
- pull and pool, the short and long u, the back of the mouth version of the first pair.
- work and walk, two vowels that merge for many learners and separate two very common verbs.
How the drill has to be built to test anything
Four conditions, and a set that misses any of them is measuring something else. The audio comes first and the word is not visible, because a learner who can see ship written down is reading rather than hearing. The choice is forced between exactly two options, because a free response lets you avoid the decision. The answer arrives immediately, since a delayed correction attaches to nothing. And the voices vary, because one speaker's version of a vowel is one speaker's version.
Most published pair sets fail the first condition, and they fail it in the most tempting way: they show you the pair while they play it, which feels helpful and converts a listening test into a reading test with sound effects. The score goes up and the ear has not moved.
Where it stops: passing the set
The stall is scoring well. A learner runs an app's pair list, gets most of them right, and concludes the sounds are learned. What actually happened is that the list was chosen by the app rather than by their language, so it contained mostly contrasts they had before they started, and the few that were genuinely theirs were a small enough share of the total that failing all of them still looked like a good score.
Nothing corrects this because the number is real. The app reports high accuracy, the learner reports progress, and the four pairs that matter are still merged. The only test that reveals it is a set built from your own language's gaps, which no general-purpose product can build for you.
Building the test yourself
Anki with the audio on the front of the card is the cheapest way to get all four conditions at once, and it is free on desktop and Android and paid once on iOS. Put the recording on the front and the written word on the back. You hear it, you commit to one of two words, you turn the card. That is a forced choice with immediate feedback and no reading.
It replaced the minimal-pair app that shows you the word while it plays the sound, which is the format nearly all of them use and which cannot test hearing at all. Two practical notes: pull recordings for the same word from several speakers rather than one, and keep the deck to the pairs your language merges, which will be a deck of a few dozen cards rather than a few hundred.
What it looks like when it goes wrong
The recognisable failure is high accuracy on a drill and no change in conversation, which is the signature of a set that tested the wrong pairs. The second is a pair that separates cleanly when the two words are played one after the other and merges when they are played apart by ten minutes, which means what was learned was the difference between two recordings rather than two categories.
The third is subtler and worth watching for: a learner who has genuinely split a pair in their hearing but still produces one sound for both. Hearing comes first and production does not follow automatically, so a drill that has worked will show up as a change in what you can hear a week or two before it shows up in what you say.
Which pairs sit close enough together to be worth checking is set out in the twenty that English hears and most languages do not and in th, r, and the handful that give every learner trouble, and the symbols for writing a pair down unambiguously are on a symbol per sound, so an unfamiliar word becomes sayable. All of it sits under being understood the first time you say it.
Once a pair has split, the open question is whether it survives outside the drill, and the honest way to find out is not another drill. It is a recording of two people talking normally, where the vowel is short, unstressed and surrounded by other sounds: real speed, real accents, no subtitles.