PhonemaBlog

Why your English app does not correct pronunciation (and what to do instead)

Most language apps check whether speech recognition understood you, which is not the same as checking whether you said the word correctly. A two-minute test tells you which kind you have.

Direct answer

Most language apps do not score your pronunciation. They check whether a speech recogniser understood you, and a recogniser is designed to succeed despite your accent. That is a useful property for a voice assistant and a useless one for a learner, because it forgives precisely the errors you wanted it to catch. How well it succeeds is also uneven: a study of five commercial recognisers published in PNAS found average word error rates of about 35% for Black speakers against 19% for white speakers of the same regions, all of them native. Read the study.

The result is the familiar experience: you say the sentence, you get a green tick, and you still have no idea whether you would be understood at a counter in London.

Recognition and pronunciation are different tests

A recogniser gets an enormous head start. It usually knows the sentence you are supposed to say, or at least knows the small set of sentences the exercise contains. It then picks the closest match. If the exercise expects I would like a ticket and you produce something roughly that shape, the match succeeds regardless of how your vowels landed.

Modern recognisers are also trained on accented speech on purpose, because handling many accents is a product requirement. That training does exactly what it should: it makes the system tolerant. Tolerance is the opposite of what a correction tool needs.

So the green tick means: a recogniser picked the expected word out of the possibilities it was given. That is a much easier test than saying the word correctly.

The two-minute test

You can find out what your app is doing without reading any marketing copy.

Step 1: pick a sentence with one difficult word

Choose a short sentence from the app that contains one word you know is hard for you, such as think, ship, or world. Keep the rest of the sentence easy.

Step 2: say it correctly and note the result

Record your best attempt and write down whatever feedback you get. This is your baseline.

Step 3: say it again with one deliberate error

Repeat the same sentence, but replace the difficult sound with the wrong one on purpose: tink for think, chip for ship. Keep everything else identical.

Step 4: compare what the app told you

If the app passed both attempts, or gave two similar numbers with no indication of where the problem was, it is checking recognition rather than pronunciation. If it pointed at the word you broke, it is doing something more useful.

Run the test twice, because a single recording can go strangely for unrelated reasons.

The four things that usually break

Scores arrive as one number per sentence. A phrase score of 72 has nowhere to put the detail. Two learners get the same 72 for opposite reasons, and neither of them learns what to change. This is why per-sound scoring exists.

Feedback arrives too late. If the answer comes back three seconds after you stop speaking, the loop is broken. The value of feedback decays fast: you need it while you still remember what your tongue did, not after your attention has moved on.

You cannot repeat freely. Apps built on server-side analysis pay per request, so they cap you: five attempts, then tomorrow. Improvement in pronunciation comes from many short repetitions of the same line, which is exactly the pattern a cap prevents.

You practise their sentences, not yours. A course sentence about a fictional trip to the beach is not the sentence you need on Monday. The words that cost you clarity are your job title, your city, your project. Those never appear in a syllabus.

What to look for instead

Four questions, and you can answer all of them from the free tier of any app.

  1. Does it tell you which word failed? Not whether the sentence passed. Which word.
  2. Can you repeat the same line as many times as you want? If there is a daily counter, the tool is priced for something other than practice.
  3. Can you use your own text? Paste a real sentence you need to say and see whether the app will work on it.
  4. How fast is the feedback? Under a second is a practice loop. Over three seconds is a quiz.

What good feedback looks like

Concretely: you say I think the third one is better, and the app tells you think and third came out unclear while everything else was fine. Now you have a decision to make instead of a number to feel bad about. You work voiceless /θ/ on its own, then in those two words, then in the whole sentence.

The same logic applies to the other classic confusions: /ʃ/ against /tʃ/ in ship and chip, English /ɹ/ in world and career, short /ɪ/ in live against the long vowel in leave.

Phonema is built around that narrower job. You bring the sentence, it tells you which word came out unclear, the analysis happens on the phone so there is no round trip and no counter. It does not teach grammar or vocabulary, and it is not trying to; those are covered better by the general apps that struggle with this one thing.

If you want the full routine rather than the diagnosis, start with how to check your English pronunciation.

Common questions

Why does my language app accept my pronunciation when it is wrong? Most apps check whether speech recognition understood the word, and recognition is built to succeed despite an accent, so it forgives exactly the errors you wanted flagged.

How can I tell if an app really scores pronunciation? Say a sentence correctly except for one word you deliberately mispronounce, and see whether the app points at that word.

Is a green tick from a speaking exercise proof my pronunciation is good? No. It is proof that a recogniser picked the expected word out of the possibilities it was given, which is a much easier test than saying the word correctly.

What should a pronunciation app show you? Which word failed, in a phrase you chose, quickly enough that you can try again while you still remember what your mouth did.

Related sounds