Why our synthetic tenor lost his high 'ee', and the one-line fix singers already use

A synthetic choir sang a tenor's high /i/ and /u/ 10-16 dB under /a/. Raising F1 to the pitch, as trained singers do, closed most of the gap.

Chant Forge turns Gregorian chant into backing tracks, scores and rehearsal files. Its preview choir is a small source-filter synthesiser: five detuned sawtooth voices through five formant resonators set to each Latin vowel, with consonants added on the same filter. The vowel targets are the textbook men's averages of Peterson and Barney (1952): /i/ has its first formant (F1) at 270 Hz, /u/ at 300 Hz.

That works for basses. It breaks for tenors. A tenor's G4 is 392 Hz and his A4 440 Hz, both above /i/'s and /u/'s first formant. Every harmonic of the note then sits on the falling side of the first resonance, and the vowel loses most of its strength. Measured from the synthesiser's own output, held 2-second notes, five random seeds each:

E4G4A4
/i/ against /a/, before-9.1 dB-16.5 dB-14.8 dB
/i/ against /a/, after-2.5 dB-3.7 dB+0.9 dB
/u/ against /a/, before-3.6 dB-11.6 dB-9.8 dB
/u/ against /a/, after-1.5 dB-1.9 dB+3.5 dB

A 16 dB drop is a note that disappears under the band every time the words land on an "i".

What singers do

Trained singers do not let this happen. On high notes they open the jaw and raise F1 so it stays at or above the pitch. Johan Sundberg calls it formant tuning (The Science of the Singing Voice, 1987). The cost is that a high /i/ drifts a little toward /e/, which is the "vowel modification" voice teachers talk about.

The fix in sing.mjs is the same rule: tuneF1 raises a vowel's F1 to at least 1.1 times the note's pitch and leaves F2, F3 and any vowel whose F1 is already higher alone. The 1.1 is our assumption (F1 just above the first harmonic). We did not fit it to recordings of singers.

Deciding before measuring

Before the first run we wrote down three lines the change had to pass, in the header of eval/f1-tune-check.mjs:

  1. at G4 and A4, /i/ and /u/ gain 2 dB or more against /a/ (it gained 12.9 dB on average);
  2. the formant reader gets no fewer held vowels right from E2 to A4 (179 of 200 before, 180 after);
  3. the corpus checks at the default key stay within half a point: vowels and final consonants came out identical, and syllable openings went from 742 to 743 of 934.

All three held, so tuning is now the default. Below B3 nothing changes, so basses sound exactly as they did.

What the measurement can't tell us

The second line is weaker than it looks. Above E4 our LPC formant reader finds F1 at about the pitch whether or not the filter was tuned: with harmonics 400 Hz apart, there is nothing between them for it to see. It reads 1 of 5 G4 /i/ notes as /i/ before the change and 1 of 5 after. So the check shows the change does no harm to what we can measure. It does not show that the tuned high /i/ sounds like /i/ to a person. Nobody has listened to it yet. The same goes for whether 1.1 is the right ratio. Both are open questions in the project's README.

Report: ventures/chant-forge/eval/f1-tune-check-report.md. Code: tuneF1 in ventures/chant-forge/sing.mjs, MIT licensed.

All writing · More in Tools and projects · More on Chant Forge · All tools and research