How does one buzz become every vowel?

In early November 2026 a voice synthesizer called CHIMERA spread across X and Hacker News. Its maker, Jameson Hodge, built it so that every sound comes from a simulated airway: air from simulated lungs, vocal folds, a throat, a mouth and a nose, with no recordings at all. Within days people had it singing barbershop quartets and Queen. It is also a very good way to see something everyone does without thinking: how one buzz becomes every vowel.

▶ Try it yourself in Vowel Lab: drag the tongue around a vowel map and hear the same buzz become “ee”, “ah” and “oo”.

The buzz is the same

When you speak a vowel, air from your lungs pushes through your vocal folds and sets them vibrating. Each vibration lets out a puff of air, about 110 times a second for an average man’s speaking voice and about 200 for a woman’s. A train of puffs is not one pitch but a whole row of them, the harmonics: 110, 220, 330 Hz and up, each a little quieter than the last. That row is the same whether you are about to say “ee”, “ah” or “oo”. Change how fast the folds vibrate and you change the note, not the vowel.

The tube picks the vowel

Above the folds, the throat bends into the mouth: a tube about 17 cm long in an adult man. Like a bottle you blow across, a tube rings at certain pitches, its resonances. Harmonics near a resonance come out loud, and the rest are damped. Phoneticians call these resonances formants and number them from the lowest: F1, F2, F3. Moving your tongue, jaw and lips reshapes the tube and moves them. With the tongue high and forward, an average man’s tube rings at about 340 and 2,300 Hz: “ee”. Drop the jaw and pull the tongue back and the peaks slide to about 770 and 1,330 Hz: “ah”. Those numbers are averages measured by James Hillenbrand and colleagues in 1995 from 45 men, 48 women and 46 children saying words like “heed” and “hod”.

Plot F1 against F2 and the vowels spread out into a map, which is why the first two formants are usually enough to tell vowels apart. Roughly, F1 rises as the jaw opens, and F2 rises as the tongue moves forward. That is the textbook rule of thumb rather than a law, and listeners also use F3, how long a vowel lasts and how it glides.

How close does CHIMERA get?

Most speech synthesizers since the 1960s work directly with formants or stitch together recordings. As Hodge put it on X, his synth “models the vocal tract as a network of branching acoustic tubes, unlike vocaloid which stitches together and pitch-shifts phoneme recordings.” It simulates the air itself, moving through a branched network of tubes from the lungs to the lips and nose, the same kind of model voice researchers use. Its vowel shapes were tuned against Hillenbrand’s averages, and by its own calibration file the formants of its 22 vowel shapes land within about 2% of their targets on average. Running speech is another matter: in the project’s own test, the Whisper speech recognizer gets about 37% of its words wrong, against 2.7% for 23 real speakers. The project’s own comparison with a real singer finds its voice source “too pressed on most notes”, and Hodge says he is “still working on improving naturalness by tuning it against recordings of real singers.” Held vowels are the easy part; consonants, timing and the way real vocal folds move are harder, and CHIMERA’s folds follow a set pattern rather than vibrating on their own.

CHIMERA is free and open source: try it in your browser or read the code.