Hear what you don't yet produce

You cannot reliably produce a sound you cannot reliably hear. Russian has no /y/ (the vowel of tu), no /ø/ (deux), and none of the four French nasal vowels — so five years in, tu and tout can still land in the same box, and blanc, blond, brin, brun blur at speed. The fix starts in the ear, not the mouth.

This drill is high-variability perception training: the same contrast spoken many times, by different voices, at different paces, in random order. You hear one word and pick which of the pair it was. That variability is the active ingredient — one careful speaker teaches you that speaker; many voices teach you the category.

Two honest notes. The voices here are your browser's French synthesis voices — provisional until recorded native clips replace them (each item is already wired to accept a real recording). And expect the first sets to feel harder than they look: that difficulty is the measurement working.

Loading the exercise...

Now say it back

Perception seeds production, but it does not guarantee it — so train the mouth too. Listen to the model tu (a different voice each time), then record yourself saying it. You get a phoneme-by-phoneme read: the /y/ is the one Russian does not have, so watch that chip.

This is the speech runtime's first spoken rung (PLT-048): your audio is scored and then discarded — nothing is stored. Scoring needs the Azure key; without it you still get the model audio to self-check against.

Now whole phrases. Read each one aloud, record, and check your score — watch the vowels Russian does not have (/y/, /ø/) and the nasals. (For these, Listen uses your browser's French voice.)