5D Library › 歌手共振峰簇、声门闭合商与声学权威度心理物理学 · English

第二章 不必喊叫就能被听见?75 dB 噪声中的共振峰簇与声门下压的代价

频带增益不等于宽带增益

「演说家共振峰簇能在 75 dB 的噪声中不必喊叫就把信噪比提高 12 dB」这一说法混淆了两个量:一个频带内的声级,与整个信号的可懂度。p03 把第一章的两个元音放进言语谱噪声中,用一个仿照言语可懂度指数的分数评估:每个倍频程带贡献其重要性乘以可听度 (American National Standards Institute 1997)。

推导

# Chapter 2. p03: What the cluster buys in 75 dB of noise. Octave-band
# levels of the two vowels of p01 (same source, scaled so the "spread"
# vowel is 65 dB overall), against speech-shaped noise of 75 dB overall
# (octave levels -2, 0, -6, -12, -18, -24 dB from 250 to 8000 Hz, a
# fall of 6 dB per octave above 500 Hz). A speech-intelligibility-index
# style score: each band contributes its octave importance (ANSI S3.5
# octave weights) times its audibility, (SNR + 15) / 30 kept in 0..1.
# "Equivalent gain": how many dB the spread voice must be raised to
# score as well as the cluster voice.
from math import log10, log2, pi, sqrt

BW = (80, 90, 120, 150, 200, 250, 300)
SPREAD = (700, 1200, 2600, 3500, 4500, 5500, 6500)
CLUSTER = (700, 1200, 2600, 2900, 3200, 5500, 6500)
CENTRES = (250, 500, 1000, 2000, 4000, 8000)
WEIGHT = (0.0617, 0.1671, 0.2373, 0.2648, 0.2142, 0.0549)
SHAPE = (-2, 0, -6, -12, -18, -24)

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 power(formants, lo, hi):
    return sum(10 ** (gain_db(f, formants) / 10)
               for f in range(int(lo), int(hi), 10))

def bands(formants, offset):
    return [10 * log10(power(formants, c / sqrt(2), c * sqrt(2))) + offset
            for c in CENTRES]

def score(speech, noise):
    return sum(w * min(max((s - n + 15) / 30, 0.0), 1.0)
               for w, s, n in zip(WEIGHT, speech, noise))

if __name__ == "__main__":
    tot = 10 * log10(power(SPREAD, 50, 11314))
    sp = bands(SPREAD, 65 - tot)
    cl = bands(CLUSTER, 65 - tot)
    ns = 10 * log10(sum(10 ** (x / 10) for x in SHAPE))
    nz = [x + 75 - ns for x in SHAPE]
    print("   band (Hz)  noise  spread  cluster   SNR spread  SNR cluster")
    for c, n, a, b in zip(CENTRES, nz, sp, cl):
        print("   {:9d}  {:5.1f}  {:6.1f}  {:7.1f}   {:10.1f}  {:11.1f}"
              .format(c, n, a, b, a - n, b - n))
    s0, s1 = score(sp, nz), score(cl, nz)
    lo, hi = 0.0, 30.0
    for _ in range(60):
        mid = (lo + hi) / 2
        if score([x + mid for x in sp], nz) < s1:
            lo = mid
        else:
            hi = mid
    print("   score: spread {:.3f}, cluster {:.3f}".format(s0, s1))
    print("   equivalent broadband gain of the cluster: {:.1f} dB".format(lo))
    print("   spread voice raised 12 dB: score {:.3f}".format(
        score([x + 12 for x in sp], nz)))
#    band (Hz)  noise  spread  cluster   SNR spread  SNR cluster
#          250   70.1    51.9     52.0        -18.1        -18.1
#          500   72.1    61.0     61.4        -11.0        -10.7
#         1000   66.1    61.7     62.4         -4.4         -3.7
#         2000   60.1    40.6     50.8        -19.5         -9.3
#         4000   54.1    29.5     47.5        -24.6         -6.5
#         8000   48.1   -17.6    -33.0        -65.7        -81.1
#    score: spread 0.106, cluster 0.225
#    equivalent broadband gain of the cluster: 6.7 dB
#    spread voice raised 12 dB: score 0.370

分散元音以 65 dB 置于 75 dB 噪声中时,分数是 0.106;聚集元音是 0.225。在 2 kHz 倍频程带,共振峰簇的信噪比是 -9.3 dB 而不是 -19.5 dB,在 4 kHz 倍频程带是 -6.5 dB 而不是 -24.6 dB;但承载大部分重要性的 1 kHz 以下频带没有变化。要得到同样的分数,分散元音需要在所有频带整体提高 6.7 dB。所以在这个模型中,共振峰簇约相当于 7 dB 的宽带声级,而不是 12 dB;把分散元音整体提高整整 12 dB,分数会是 0.370,远高于共振峰簇。

两个元音在 75 dB 言语谱噪声中的倍频程带信噪比(p03)。

被听见的代价

在噪声中,说话者会不自觉地提高嗓门,这就是伦巴德效应 (Lane and Tranel 1971)。提高嗓门要付出声门下压的代价;声压每加倍,声级约增加 8 至 9 dB (Titze and Sundberg 1992)。p04 把每 dB 噪声提高 0.5 dB 的伦巴德斜率与 p03 的 6.7 dB 结合起来。

推导

# Chapter 2. p04: The price of being heard, in subglottal pressure. A
# speaker who raises the voice by 0.5 dB per dB of noise above 50 dB (a
# Lombard slope; reported slopes range from about 0.3 to 0.6) starts
# from 0.6 kPa of subglottal pressure in quiet. Sound level rises by
# 8 to 9 dB per doubling of pressure, so a rise of R dB needs pressure
# 0.6 * 2^(R / slope). With a cluster worth G = 6.7 dB (p03), the rise
# needed for the same intelligibility is R - G (not below 0).
G, P0, LOMBARD = 6.7, 0.6, 0.5

def pressure(rise, slope):
    return P0 * 2 ** (max(rise, 0.0) / slope)

if __name__ == "__main__":
    print("   noise  rise   Ps (kPa) 8 dB  9 dB   with cluster: rise"
          "  8 dB  9 dB")
    for n in (55, 65, 75, 85):
        r = LOMBARD * max(n - 50, 0)
        fmt = ("   {:5d}  {:4.1f}  {:13.2f}  {:4.2f}  {:19.1f}  {:4.2f}"
               "  {:4.2f}")
        print(fmt.format(n, r, pressure(r, 8), pressure(r, 9),
                      max(r - G, 0.0), pressure(r - G, 8),
                      pressure(r - G, 9)))
    print("   pressure saved at 75 dB of noise: {:.0%} (8 dB) to {:.0%}"
          " (9 dB)".format(1 - 2 ** (-G / 8), 1 - 2 ** (-G / 9)))
#    noise  rise   Ps (kPa) 8 dB  9 dB   with cluster: rise  8 dB  9 dB
#       55   2.5           0.75  0.73                  0.0  0.60  0.60
#       65   7.5           1.15  1.07                  0.8  0.64  0.64
#       75  12.5           1.77  1.57                  5.8  0.99  0.94
#       85  17.5           2.73  2.31                 10.8  1.53  1.38
#    pressure saved at 75 dB of noise: 44% (8 dB) to 40% (9 dB)

在 75 dB 噪声中,伦巴德效应使声级提高 12.5 dB;按每加倍 8 dB 计需要 1.77 kPa 的声门下压,按 9 dB 计需要 1.57 kPa,而安静时只需 0.6 kPa。有了共振峰簇,所需的提高降到 5.8 dB,声压降到 0.99 或 0.94 kPa:节省约 40% 至 44%。共振峰簇让说话者「不必喊叫」就能被听见,正是在这个意义上:不是多加 12 dB,而是让每一分力气在要紧的频带里发挥更大的作用。

假说 · 尚未验证

作者的诠释(尚未验证): 经过缩小 F3 至 F5 间距的训练后,在 75 dB 噪声的房间里讲话的人,为达到同样的可懂度所用的声门下压(由清塞音期间的口内压估计)低于训练前。什么会证明它错: 在一项预先登记、测量口内压的训练研究中,噪声中按可懂度匹配的声压在训练后没有下降,或在只训练响度的对照组中下降得同样多。

小结

在模型中,共振峰簇在 2 至 4 kHz 频带的 14 dB,在 75 dB 言语谱噪声中对可懂度约相当于 7 dB 的宽带声级。它真正的好处是省力:在那样的噪声中被听懂所需的声门下压约减少五分之二。

参考文献

American National Standards Institute. 1997. Methods for Calculation of the Speech Intelligibility Index (ANSI S3.5-1997). Acoustical Society of America.
Lane, Harlan, and Bernard Tranel. 1971. “The Lombard Sign and the Role of Hearing in Speech.” Journal of Speech and Hearing Research 14 (4): 677 to 709.
Titze, Ingo R., and Johan Sundberg. 1992. “Vocal Intensity in Speakers and Singers.” The Journal of the Acoustical Society of America 91 (5): 2936 to 2946.

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).