Where it landed
Hold a steady vowel to see its position.
The marker follows your voice while you hold the vowel. Positions come from formants; the frame is the IPA schematic of height and backness, and formants cannot recover where the tongue actually was. The marker and the verdict are measured separately, so they can disagree.
Practising a particular vowel? Optional — the model is given your audio and nothing else, so it names the sound either way. Pick one only to be told whether you hit it, and click it again to stop aiming.
Fetching the model — this takes about half a minute
Starting…
It is 203 MB, because the whole thing runs on your machine rather than on a server. Your browser keeps it, so this only happens once — every later visit starts instantly. A fuller, slightly more accurate model downloads quietly afterwards and takes over by itself.
Press Start listening and hold any vowel — it will say what it hears.
I was actually saying
A clip is only worth having if it comes with the sound you meant, so say which one — a clip the model got wrong is the most useful kind. Donated clips are stored publicly under CC BY-SA 4.0 and used to train and test this model. Do not donate anything you would not publish. Nothing else is ever uploaded.
It is not very good yet. That is the point.
On 81 held vowels from British teaching channels the model had never heard, it gets 76 right. Measured end to end through this page rather than by feeding the model directly, the quick model it starts on scores the same — 77 of the 81 — so on this evidence the two are hard to tell apart in use. Those clips were checked one by one by ear, and the model that produces them is trained only on OpenSLR SLR83 (CC BY-SA 4.0), so the weights can be shared.
Known limits, stated plainly: only the five long vowels work — the short ones (bed, cat, cup) are not offered because the segmenter that builds training data cannot isolate them yet. The test set leans on two speakers. Every verdict is experimental, and if you are not a southern British speaker it will disagree with you in ways that may be its fault, not yours.
If it gets your vowel wrong, that is useful. Donate the clip and say so.