My Upsilon Problem

Before graduation, all my Classics instructors encouraged continued reading and study. Otherwise, we would lose a lot of what we learned. And I have.

The Greek alphabet came back quickly, which was reassuring. I expected the first few days to be mostly priming and retrieval: letters, familiar endings, the shape of words I once knew.

Then I tried to say ψυχή.

In scholarly transliteration, it is psychē. The reconstructed fifth-century Attic pronunciation is /psyː.kʰɛ̌ː/. It is decidedly NOT the English psyche, “sigh-key.”

The problem is I cannot properly pronounce upsilon because hearing and pronunciation are skills you can develop (and lose). As English does not have /y/ as a phoneme, the simplest way to make /y/ was to hold the tongue position for the vowel in feed—/i/—and round my lips. I had to start with the vowels I do have. /i/ is high and front with unrounded lips; /y/ is high and front with rounded lips; /u/ is high and back with rounded lips. The lips distinguish /i/ from /y/; the tongue position distinguishes /y/ from /u/.

I am not particularly keen on phonetics as a field. Apparently, though, I am keen on phonetics when it turns into a pesky enough problem that I can solve through terrible vocalization attempts. (DO NOT LAUGH AT MY BAD DATA.) Twenty minutes later, I was comparing formants across ten-plus recordings. By the time I stopped, I had thirteen (with various degrees of success).

The measurements are useful because ears are highly suggestible. Very roughly, F1 tends to move with vowel height, while F2 tends to move with how front or back the tongue is. A front vowel such as /i/ usually has a higher F2 than a back vowel such as /u/; lip rounding generally lowers their resonances. Vocal tracts differ, recording conditions differ, and formant trackers make mistakes. My useful evidence came from comparing my own /i/, intended /y/, and /u/ within the same take, while paying attention to signal quality.

The workflow did not make the interpretations automatic or infallible. Several early readings were plain wrong. I had not labeled the spoken sequences, so automatic segmentation sometimes treated speech, pauses, and the wrong vowels as the tokens I meant to compare. Background noise confused detection.

Recording 11 was my breakthrough: the cleanest stable isolated /y/, with F2 mostly around 2015–2290 Hz and F1 around 340–386 Hz. It finally sounded and measured like I had kept the tongue front while rounding instead of sliding all the way back to /u/.

Recording 13 was more useful still because it contained success and failure under nearly the same conditions. In the first /i–y–u/ sequence, the contrast is audible and visible: the first /i/ sat around 2800–2870 Hz in F2, the first /y/ around 2130–2490 Hz, and /u/ around 700–1000 Hz. Later in the same take, a second /i–y/ was weaker. The second /y/, around 2500–2640 Hz, stayed too close to the preceding /i/.

All of this was just to say ψυχή a little more honestly: not “sigh-key,” and not with upsilon quietly collapsing into either /i/ or /u/. The disproportion is ridiculous, which is part of why it was fun. I wanted one rounded vowel in one familiar word and accidentally built a mini lab experiment.

Materials

• Surface Laptop built-in microphone
• GitHub Copilot
• Primary Copilot chat model: GPT-5.6 Sol