5D Library › Singer Formant Clustering, Glottal Closure Quotients and Acoustic Authority Psychophysics · 中文

Chapter 1. The Speaker’s Formant Cluster: What Moving F3, F4 and F5 Together Does to a Spectrum

The source-filter picture

A voiced sound is a glottal source shaped by the resonances of the vocal tract, the formants (Fant 1960). Trained classical singers show a strong spectral peak near 3 kHz, the “singer’s formant”, which Sundberg explained as a clustering of the third, fourth and fifth formants produced by a widened pharynx and a narrow outlet of the larynx tube (Sundberg 1974). A weaker counterpart in the speaking voices of actors and trained speakers, near 3500 Hz, is called the speaker’s or actor’s formant (Nawka et al. 1997). p01 builds the envelope of a vowel from seven resonators and compares an ordinary spread of F3, F4 and F5 with a cluster.

Calculation

# Chapter 1. p01: The speaker's formant cluster in a source-filter model.
# Seven cascaded formant resonators (vowel /a/: F1 700, F2 1200 Hz; F6
# 5500 and F7 6500 Hz fixed) on a source-plus-radiation tilt of -6 dB
# per octave above 100 Hz. "Spread" places F3, F4, F5 at 2600, 3500,
# 4500 Hz (ordinary speech); "cluster" at 2600, 2900, 3200 Hz.
# Bandwidths 80, 90, 120, 150, 200, 250, 300 Hz. Levels are power sums
# on a 10 Hz grid from 50 to 8000 Hz, in dB.
from math import log10, log2, pi

BW = (80, 90, 120, 150, 200, 250, 300)
SPREAD = (700, 1200, 2600, 3500, 4500, 5500, 6500)
CLUSTER = (700, 1200, 2600, 2900, 3200, 5500, 6500)

def gain_db(f, formants):
    s = 2j * pi * f
    g = 1.0
    for fn, bn in zip(formants, BW):
        p = complex(-pi * bn, 2 * pi * fn)
        g *= abs(p) ** 2 / (abs(s - p) * abs(s - p.conjugate()))
    tilt = -6 * log2(f / 100) if f > 100 else 0.0
    return 20 * log10(g) + tilt

def band(formants, lo, hi):
    tot = sum(10 ** (gain_db(f, formants) / 10)
              for f in range(lo, hi, 10))
    return 10 * log10(tot)

if __name__ == "__main__":
    print("   band (Hz)        spread   cluster   difference")
    for lo, hi in ((50, 8000), (50, 1000), (1000, 2000), (2000, 4000),
                   (2500, 3500), (4000, 8000)):
        a, b = band(SPREAD, lo, hi), band(CLUSTER, lo, hi)
        print("   {:>4} to {:<4}  {:8.1f}  {:8.1f}  {:+10.1f}".format(
            lo, hi, a, b, b - a))
    for f in (2600, 2900, 3200):
        print("   at {} Hz: cluster {:+.1f} dB relative to spread".format(
            f, gain_db(f, CLUSTER) - gain_db(f, SPREAD)))
#    band (Hz)        spread   cluster   difference
#      50 to 8000      19.6      20.2        +0.6
#      50 to 1000      18.9      19.2        +0.3
#    1000 to 2000      11.2      12.4        +1.2
#    2000 to 4000      -7.4       6.8       +14.2
#    2500 to 3500      -8.7       6.6       +15.2
#    4000 to 8000     -31.6     -55.8       -24.2
#    at 2600 Hz: cluster +12.7 dB relative to spread
#    at 2900 Hz: cluster +25.6 dB relative to spread
#    at 3200 Hz: cluster +15.5 dB relative to spread

Moving F4 and F5 down next to F3 adds almost nothing to the overall level (+0.6 dB) or to the band below 1 kHz (+0.3 dB) that carries most of the energy. In the band from 2 to 4 kHz the gain is +14.2 dB, and +15.2 dB between 2500 and 3500 Hz; at 2900 Hz, the centre of the cluster, the envelope is 25.6 dB higher. Above 4 kHz the cluster loses 24.2 dB, because the resonances that used to sit there have moved down. The cluster does not make the voice louder; it moves energy into the region where the ear is most sensitive and where intelligibility is decided.

Envelope of the “spread” and “cluster” vowels of p01; the shaded band is 2 to 4 kHz.

How close is close

p02 keeps F4 at 2900 Hz and varies the spacing to F3 and F5.

Calculation

# Chapter 1. p02: How close must F3, F4 and F5 be? The model of p01 with
# F4 at 2900 Hz and F3, F5 at 2900 minus and plus a spacing d. Gain in
# the 2000 to 4000 Hz band relative to the "spread" vowel of p01, and
# the change of the overall level. A singer's or actor's cluster is
# usually described with spacings of a few hundred hertz.
from math import log10, log2, pi

BW = (80, 90, 120, 150, 200, 250, 300)
SPREAD = (700, 1200, 2600, 3500, 4500, 5500, 6500)

def gain_db(f, formants):
    s = 2j * pi * f
    g = 1.0
    for fn, bn in zip(formants, BW):
        p = complex(-pi * bn, 2 * pi * fn)
        g *= abs(p) ** 2 / (abs(s - p) * abs(s - p.conjugate()))
    return 20 * log10(g) + (-6 * log2(f / 100) if f > 100 else 0.0)

def band(formants, lo, hi):
    return 10 * log10(sum(10 ** (gain_db(f, formants) / 10)
                          for f in range(lo, hi, 10)))

if __name__ == "__main__":
    ref_b, ref_t = band(SPREAD, 2000, 4000), band(SPREAD, 50, 8000)
    print("   spacing (Hz)  F3-F4-F5            2-4 kHz gain  overall")
    for d in (900, 700, 500, 400, 300, 200, 150):
        fs = (700, 1200, 2900 - d, 2900, 2900 + d, 5500, 6500)
        print("   {:12d}  {:>4}-{:>4}-{:<4}      {:+12.1f}  {:+7.1f}".format(
            d, fs[2], fs[3], fs[4], band(fs, 2000, 4000) - ref_b,
            band(fs, 50, 8000) - ref_t))
#    spacing (Hz)  F3-F4-F5            2-4 kHz gain  overall
#             900  2000-2900-3800              +3.3     +1.0
#             700  2200-2900-3600              +6.3     +0.7
#             500  2400-2900-3400              +8.5     +0.6
#             400  2500-2900-3300             +10.8     +0.6
#             300  2600-2900-3200             +14.2     +0.6
#             200  2700-2900-3100             +19.2     +1.0
#             150  2750-2900-3050             +22.4     +1.6

With 900 Hz spacing the band gain is only +3.3 dB; at 500 Hz it is +8.5 dB, at 300 Hz +14.2 dB and at 150 Hz +22.4 dB. The gain grows steeply as the formants approach each other, because the peaks of neighbouring resonances add. In this model a band gain of about 12 dB needs a spacing between 300 and 400 Hz. Real formant bandwidths widen when formants crowd together, so the steepest part of this curve is an upper bound.

Hypothesis · not established

The author’s interpretation (not yet tested): speakers judged “resonant” or “carrying” by listeners have a smaller average spacing between F3, F4 and F5 in their long-term spectra than speakers of the same sex judged “thin”, independent of overall loudness. What would show it wrong: in a corpus of read speech with listener ratings, the F3 to F5 spacing estimated from the long-term average spectrum is unrelated to “carrying” ratings once speech level is controlled.

Summary

Clustering F3, F4 and F5 adds about 14 dB in the 2 to 4 kHz band of the model while leaving the overall level almost unchanged; the gain depends steeply on how close the formants come. It is a redistribution of energy, not extra effort.

References

Fant, Gunnar. 1960. Acoustic Theory of Speech Production. Mouton.
Nawka, Tadeus, Lutz-Christian Anders, Mario Cebulla, and David Zurakowski. 1997. “The Speaker’s Formant in Male Voices.” Journal of Voice 11 (4): 422 to 428.
Sundberg, Johan. 1974. “Articulatory Interpretation of the “Singing Formant”.” The Journal of the Acoustical Society of America 55 (4): 838 to 844.

Fan, Y. (2026). Singer Formant Clustering, Glottal Closure Quotients and Acoustic Authority Psychophysics: Subglottal Pressure Dynamics, Pitch-Floor Stability and Pause-to-Phoneme Persuasion Ratios [歌手共振峰簇、声门闭合商与声学权威度心理物理学] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.24090258 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).