Where does your accent live? Everywhere.

We ran the experiment everyone asks for. People want a shortcut: "If I just fix my vowels, will 80% of my accent disappear?" So we took 1,694 recordings of people from 29 language backgrounds reading the same paragraph, and started deleting sounds — surgically, class by class — to see what the AI actually needs to place an accent.

The experiment: silencing sounds, one family at a time

Every recording was aligned to the phoneme level, so we knew exactly where each vowel, each consonant, each r and l lived on the timeline. Then we silenced entire families of sounds in the audio itself — every vowel gone, or every consonant, or every liquid — and asked the model to identify the speaker's native language from what remained.

The answer nobody expected

Delete every vowel a person says, and the AI still identifies their native language almost as well as before — from consonants alone. Delete every consonant instead, and the vowels alone carry it. Remove just the r's and l's, just the stops, just the nasals: the needle barely moves. And timing alone — the rhythm of the segments with all the sound taken away — identifies nothing above chance.

Your accent is not hiding in a sound. It is a whole-system signature, written redundantly into everything you say: every vowel shade, every consonant release, all at once. The evidence is so spread out that no single deletion can erase it — like a watermark printed on every page of the book, not just the cover.

But the sounds are not equal for everyone

The averages hide the useful part. When we deleted consonants, speakers of some languages became dramatically harder to identify — and the pattern matches what phonologists would predict:

• Urdu speakers lost the most (−21 recall): the retroflex consonants are their signature's loudest thread.
• The Slavic and Germanic groups (Russian, Serbian, Bulgarian, Polish, German, Swedish) each lost 8–11: think final-consonant devoicing and th-handling.
• Italian speakers, by contrast, were hurt most when we deleted vowels — the most vowel-led accent in the study.

So the honest coaching advice isn't "fix your vowels." It's: which sounds matter most depends on your first language — and now that ordering is measured, not guessed.

What this means if you're trying to soften an accent

There is no one weird trick. A machine — and to a large degree, an attentive ear — reads your accent from the whole fabric of your speech. That's genuinely good news for learners: it means accent work isn't about policing one embarrassing sound; it's about gradually retuning the whole system, starting with the threads that matter most for your language background. That's exactly the philosophy behind Mimica, the tutor this research feeds: whole conversations, in your own voice, prioritized by what your first language actually needs.

Curious what the whole-system signature says about you?

take the 20-second accent test →