Why pronunciation stops improving, and what NoaLingua does about it

The plateau is in your ear, not your mouth

Everyone notices the same thing eventually: after some months of speaking, pronunciation stops getting better. More conversation does not move it. More listening does not move it. It sits where it is.

The reason is that perception comes first. Your first language taught your ear to sort sounds into categories and to discard the differences inside a category, because those differences never mattered. When a new language uses one of those discarded differences to distinguish two words, your ear files both under the same heading and reports back that they sound identical.

That is why the plateau feels like a physical limit. You cannot correct an error you cannot hear, and no amount of practice fixes a target you are not perceiving.

What actually moves it

Three things, in order of how much they return.

Immediate comparison, same sentence

Say a line, then hear the original again straight away. The gap between the two is the only feedback that is specific enough to act on, and it disappears from memory within seconds.

Repetition on the same line, not new material

Improvement happens across repetitions of one sentence. Moving on to a new sentence each time exercises fluency and leaves pronunciation exactly where it was.

Prosody before phonemes

Rhythm and intonation carry more of the impression of an accent than individual sounds do. A sentence with correct stress and wrong vowels is understood; the reverse often is not.

Shadowing is the mechanism, and its problem is mechanical

Shadowing — repeating a line aloud immediately after hearing it, in the same rhythm — does all three of those at once. It was developed for interpreter training and it works for exactly the reason above: you cannot shadow a sentence you did not parse, so it forces perception and production together.

The difficulty is not intellectual, it is mechanical. Doing it properly means playing a line, pausing, repeating, rewinding, and doing that forty times in a session. Most people give up on the scrub bar long before they give up on the technique, which is why automating the loop matters more than it sounds: the sentence plays, the video stops itself, and the pause is deliberately longer than the line, because producing a foreign sentence takes longer than hearing one.

What a pronunciation score can honestly tell you

Speech recognition can tell you one thing well: whether a listener who is not trying to understand you would have got the words. That is a real and useful signal, and it is not the same as an accent rating.

What it cannot do is judge how native you sound, because it was never built for that. It also mishears names, it is worse on short utterances than long ones, and it is more forgiving of a strong accent than a human listener is. Any tool that turns that into a single percentage and calls it your accent is overreading its own data.

The honest version is to report what was heard against what was said, name the words that did not come through, and stop there. That is what Say It Back does, and the limits page says plainly what the score does not mean. Nothing is recorded: the score and the missed words are kept, the audio never is.

A session that works

Fifteen minutes, on material you have already understood.

1. Pick a clip you have already watched for meaning. Comprehension and production are separate passes; doing both at once does neither well.
2. Shadow the whole thing once without stopping, to find the lines that fall apart.
3. Take three of those lines and drill each one five times, listening to the original between attempts.
4. Score the three at the end. If the same word fails every time, that is a perception problem — listen to it in isolation before trying again.
5. Stop. Pronunciation practice past about twenty minutes stops returning anything, because the muscles and the attention both tire.