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
第五章 从指标到评分:以信任评分对嗓音做回归能还原什么
一项合成的评分研究
听众确实会对嗓音作出反应:例如在男女被试的实验中,音高较低的嗓音都更被看好担任领导 (Klofstad et al. 2012)。要核对的是一个更强的说法:多元回归能给出基频微扰、振幅微扰、句末降调与停顿占比对信任与支配感的可靠「权重」。p09 模拟一项包含 120 段评分录音、真实权重已知的研究。
推导
# Chapter 5. p09: What a regression of "trust" ratings on voice measures
# can recover. 120 synthetic recordings, each rated for trust (mean of
# a listener panel, 1 to 7). Measures: mean F0 (st above 100 Hz),
# jitter %, shimmer % (correlated with jitter, r about 0.7), final slope
# (st/s; more negative = stronger fall) and pause ratio (correlated with
# the fall, r about 0.5). True model: trust = 4 - 0.10 F0 - 0.30 jitter
# - 0.05 shimmer - 0.03 slope - 30 (pause - 0.21)^2 + noise (sd 0.5);
# the square expands to 12.6 pause - 30 pause^2 plus a constant.
# Ordinary least squares with pause and pause^2; standard errors and
# variance inflation factors (VIF). Seed 99.
import random
from math import sqrt
def solve(a, b):
n = len(a)
m = [row[:] + [v] for row, v in zip(a, b)]
for i in range(n):
p = max(range(i, n), key=lambda r: abs(m[r][i]))
m[i], m[p] = m[p], m[i]
for r in range(n):
if r != i:
f = m[r][i] / m[i][i]
m[r] = [x - f * y for x, y in zip(m[r], m[i])]
return [m[i][n] / m[i][i] for i in range(n)]
def ols(rows, y):
k = len(rows[0])
xtx = [[sum(r[i] * r[j] for r in rows) for j in range(k)]
for i in range(k)]
beta = solve(xtx, [sum(r[i] * v for r, v in zip(rows, y))
for i in range(k)])
res = [v - sum(b * x for b, x in zip(beta, r)) for r, v in zip(rows, y)]
s2 = sum(e * e for e in res) / (len(y) - k)
inv = [solve(xtx, [1.0 if i == j else 0.0 for i in range(k)])
for j in range(k)]
return beta, [sqrt(s2 * inv[i][i]) for i in range(k)], res
def r2(rows, y):
_, _, res = ols(rows, y)
m = sum(y) / len(y)
return 1 - sum(e * e for e in res) / sum((v - m) ** 2 for v in y)
NAMES = ("F0 (st)", "jitter %", "shimmer %", "final slope", "pause",
"pause^2")
TRUE = (-0.10, -0.30, -0.05, -0.03, 12.6, -30.0)
if __name__ == "__main__":
rng = random.Random(99)
rows, y = [], []
for _ in range(120):
f0 = rng.gauss(2, 2)
z = rng.gauss(0, 1)
jit = 0.6 + 0.2 * z
shim = 3.0 + 0.8 * (0.7 * z + 0.714 * rng.gauss(0, 1))
w = rng.gauss(0, 1)
slope = -8 + 6 * w
pause = 0.20 + 0.06 * (-0.5 * w + 0.866 * rng.gauss(0, 1))
t = (4 - 0.10 * f0 - 0.30 * jit - 0.05 * shim - 0.03 * slope
- 30 * (pause - 0.21) ** 2 + rng.gauss(0, 0.5))
rows.append([1.0, f0, jit, shim, slope, pause, pause * pause])
y.append(t)
beta, se, _ = ols(rows, y)
print(" measure true estimate s.e. t VIF")
for i, name in enumerate(NAMES, 1):
others = [[r[0]] + [r[j] for j in range(1, 7) if j != i]
for r in rows]
vif = 1 / (1 - r2(others, [r[i] for r in rows]))
tv = "{:6.2f}".format(TRUE[i - 1])
print(" {:<12} {} {:9.3f} {:6.3f} {:5.1f} {:7.1f}".format(
name, tv, beta[i], se[i], beta[i] / se[i], vif))
print(" model R^2 {:.3f}".format(r2(rows, y)))
# measure true estimate s.e. t VIF
# F0 (st) -0.10 -0.110 0.023 -4.8 1.0
# jitter % -0.30 -0.123 0.318 -0.4 2.4
# shimmer % -0.05 -0.048 0.084 -0.6 2.5
# final slope -0.03 -0.038 0.009 -4.4 1.4
# pause 12.60 14.176 3.318 4.3 20.3
# pause^2 -30.00 -33.448 8.243 -4.1 20.1
# model R^2 0.386回归能还原相对其离散程度较大的效应:平均音高(-0.110,真实值 -0.10,t 值 -4.8)与句末斜率(-0.038,真实值 -0.03)。它还原不了基频微扰:估计值是 -0.123,标准误 0.318,而真实权重是 -0.30,因为基频微扰与振幅微扰同进同退(方差膨胀因子 2.4 与 2.5),且在健康样本中变化很小。停顿项被还原了,方差膨胀因子约为 20,这在一个变量与其平方同时进入模型时是必然的。模型解释了 0.386 的方差;这类研究发表的「权重」,既是关于听众的陈述,也同样是关于该样本中嗓音分布的陈述。
最佳停顿占比
p10 问一项研究能否把倒 U 形的最优点定位到流行的 0.18 至 0.25 区间内,这里真实最优点是 0.21。
推导
# Chapter 5. p10: Can a study pin down the "best" pause ratio? Trust =
# a - c (pause - 0.21)^2 + noise (sd 0.6), with pause ratios spread as
# in p09 (mean 0.20, sd 0.06). Each simulated study has N recordings,
# fits a + b1 pause + b2 pause^2 by least squares and reports the
# optimum -b1 / (2 b2). Curvature c = 30 (as in p09) or 10 (a flatter
# hill). 2,000 studies per case, seed 99.
import random
def fit(xs, ys):
n = len(xs)
s = [sum(x ** k for x in xs) for k in range(5)]
t = [sum(y * x ** k for x, y in zip(xs, ys)) for k in range(3)]
a = [[s[0], s[1], s[2]], [s[1], s[2], s[3]], [s[2], s[3], s[4]]]
for i in range(3):
for r in range(3):
if r != i:
f = a[r][i] / a[i][i]
a[r] = [u - f * v for u, v in zip(a[r], a[i])]
t[r] -= f * t[i]
return [t[i] / a[i][i] for i in range(3)] if n else None
if __name__ == "__main__":
rng = random.Random(99)
print(" curvature N optimum 10% median 90% in 0.18-0.25"
" no peak")
for c in (30, 10):
for n in (60, 120, 600):
opts, nopeak = [], 0
for _ in range(2000):
xs = [rng.gauss(0.20, 0.06) for _ in range(n)]
ys = [4 - c * (x - 0.21) ** 2 + rng.gauss(0, 0.6)
for x in xs]
_, b1, b2 = fit(xs, ys)
if b2 >= 0:
nopeak += 1
else:
opts.append(-b1 / (2 * b2))
opts.sort()
q = lambda a: opts[int(a * (len(opts) - 1))]
print(" {:9d} {:4d} {:11.3f} {:7.3f} {:5.3f} {:12.3f}"
" {:7.3f}".format(c, n, q(0.1), q(0.5), q(0.9),
sum(0.18 <= o <= 0.25 for o in opts)
/ 2000, nopeak / 2000))
# curvature N optimum 10% median 90% in 0.18-0.25 no peak
# 30 60 0.179 0.209 0.247 0.767 0.044
# 30 120 0.189 0.210 0.237 0.905 0.004
# 30 600 0.201 0.210 0.220 1.000 0.000
# 10 60 0.139 0.209 0.301 0.402 0.266
# 10 120 0.152 0.207 0.290 0.483 0.185
# 10 600 0.183 0.209 0.245 0.809 0.022山峰明显(曲率 30)且有 60 段录音时,十项研究中有八项估计的最优点落在 0.179 至 0.247 之间,0.767 的研究落在区间内;600 段录音时全部落在区间内。山峰较平(曲率 10)时,60 段录音把最优点定在 0.139 至 0.301 之间,只有 0.402 的机会落在区间内,另有 0.266 的研究根本找不到峰。即使 600 段录音也只有 0.809。一个尖锐的「18% 至 25%」区间,所宣称的精度超过了大多数评分研究所能提供的。
假说 · 尚未验证
作者的诠释(尚未验证): 在有说服力的言语评分研究中,估计出的最佳停顿占比在不同研究间的稳定性,不如平均音高效应的估计。什么会证明它错: 在至少五项同时报告两者的独立评分研究的元分析中,停顿最优点相对其标准误的研究间离散,不大于音高效应的离散。
小结
以评分对嗓音指标做回归,能还原平均音高这类大而独立的效应,却还原不了基频微扰与振幅微扰这类小而相关的效应。倒 U 形的最优点需要数百段录音才能定位到 0.18 至 0.25 这样窄的区间之内。
参考文献
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).