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
第八章 实践:权威嗓音诊断的测量规程
注
不是医疗或临床工具。 本章概述如何为辅导与讨论收集嗓音指标。基频微扰、振幅微扰与闭合商也用于临床嗓音评估;声音嘶哑持续数周以上应请医生检查。本章的参数均为示意。
哪些指标可以重复
一个为每人报告一个数字的诊断,首先要证明这个数字逐日稳定。p14 用示意的人际与个人内标准差计算信度(组内相关系数 (Shrout and Fleiss 1979))与最小可检测变化。
推导
# Chapter 8. p14: How repeatable are the measures? Illustrative between-
# person and within-person (day-to-day) standard deviations. Reliability
# ICC = b^2 / (b^2 + w^2); the smallest change in one person that is
# larger than day-to-day noise with 95% confidence is MDC = 1.96 *
# sqrt(2) * w. A study of 40 persons measured on two days estimates the
# ICC; its 10% to 90% range over 2,000 such studies is shown. Seed 99.
import random
from math import sqrt
MEASURES = (("mean F0 (st)", 2.5, 0.8), ("jitter %", 0.15, 0.12),
("shimmer %", 0.8, 0.7), ("pause ratio", 0.05, 0.04),
("final slope st/s", 5.0, 6.0))
def icc(pairs):
n = len(pairs)
m = sum(a + b for a, b in pairs) / (2 * n)
msb = 2 * sum(((a + b) / 2 - m) ** 2 for a, b in pairs) / (n - 1)
msw = sum((a - b) ** 2 / 2 for a, b in pairs) / n
return (msb - msw) / (msb + msw)
if __name__ == "__main__":
rng = random.Random(99)
print(" measure ICC MDC study ICC 10% to 90%")
for name, b, w in MEASURES:
est = []
for _ in range(2000):
pairs = []
for _ in range(40):
t = rng.gauss(0, b)
pairs.append((t + rng.gauss(0, w), t + rng.gauss(0, w)))
est.append(icc(pairs))
est.sort()
print(" {:<16} {:5.2f} {:6.2f} {:6.2f} to {:5.2f}".format(
name, b * b / (b * b + w * w), 1.96 * sqrt(2) * w,
est[200], est[1800]))
# measure ICC MDC study ICC 10% to 90%
# mean F0 (st) 0.91 2.22 0.87 to 0.94
# jitter % 0.61 0.33 0.46 to 0.72
# shimmer % 0.57 1.94 0.40 to 0.69
# pause ratio 0.61 0.11 0.47 to 0.72
# final slope st/s 0.41 16.63 0.23 to 0.57平均音高的重复性很高(0.91),最小可检测变化是 2.22 个半音。基频微扰与停顿占比中等(0.61),振幅微扰较低(0.57),句末斜率很差(0.41):单日的斜率要变化 16.63 st/s 才能超过逐日噪声。40 人的研究对这些信度的估计范围很宽(斜率为 0.23 至 0.57),所以诊断工具应公布自己的重测数据。
麦克风放在哪里
p15 把一个在 0.3 米处为 70 dB 的嗓音放进一间典型的会议室。
推导
# Chapter 8. p15: Where to put the microphone. A voice of 70 dB at 0.3 m
# (direct sound falls 6 dB per doubling of distance) in a room of 100
# m^3 with reverberation time 0.6 s and background noise of 45 dB. The
# critical distance, where direct and reverberant sound are equal, is
# 0.057 sqrt(Q V / T) with directivity Q = 2. Shown: direct level,
# direct-to-reverberant ratio (D/R) and signal-to-noise ratio (SNR). The
# rule of thumb used in this book: D/R of at least 10 dB and SNR of at
# least 30 dB for perturbation measures.
from math import log10, sqrt
V, T60, Q, NOISE = 100.0, 0.6, 2.0, 45.0
if __name__ == "__main__":
rc = 0.057 * sqrt(Q * V / T60)
print(" critical distance {:.2f} m".format(rc))
print(" distance (m) direct (dB) D/R (dB) SNR (dB) usable")
for r in (0.05, 0.1, 0.2, 0.3, 0.5, 1.0, 2.0):
lvl = 70 - 20 * log10(r / 0.3)
dr = 20 * log10(rc / r)
snr = lvl - NOISE
print(" {:12.2f} {:11.1f} {:8.1f} {:8.1f} {:>6}".format(
r, lvl, dr, snr, "yes" if dr >= 10 and snr >= 30 else "no"))
# critical distance 1.04 m
# distance (m) direct (dB) D/R (dB) SNR (dB) usable
# 0.05 85.6 26.4 40.6 yes
# 0.10 79.5 20.3 34.5 yes
# 0.20 73.5 14.3 28.5 no
# 0.30 70.0 10.8 25.0 no
# 0.50 65.6 6.4 20.6 no
# 1.00 59.5 0.3 14.5 no
# 2.00 53.5 -5.7 8.5 no临界距离是 1.04 米;在 1 米处直达声与混响声相当。按本书的经验法则(直达混响比至少 10 dB、信噪比至少 30 dB),只有 0.05 与 0.10 米的距离合格;0.3 米处的信噪比是 25.0 dB。距离固定的头戴麦克风是实际可行的办法。
基线需要多少天
p16 把 p14 的逐日变化换算成达到某一精度的个人基线所需的录音天数。
推导
# Chapter 8. p16: How many recording days does a personal baseline need?
# With the day-to-day standard deviations w of p14, the mean of k days
# has a 95% half-width of 1.96 w / sqrt(k). Shown for k = 1 to 20 days,
# and the days needed for a chosen precision.
from math import ceil, sqrt
MEASURES = (("mean F0 (st)", 0.8, 0.5), ("jitter %", 0.12, 0.05),
("shimmer %", 0.7, 0.3), ("pause ratio", 0.04, 0.02),
("final slope st/s", 6.0, 3.0))
if __name__ == "__main__":
ks = (1, 3, 5, 10, 20)
print(" measure " + "".join(" k={:<4d}".format(k)
for k in ks) + " target days")
for name, w, target in MEASURES:
row = "".join(" {:6.3f}".format(1.96 * w / sqrt(k)) for k in ks)
need = ceil((1.96 * w / target) ** 2)
print(" {:<16} {} {:6.2f} {:4d}".format(name, row, target, need))
# measure k=1 k=3 k=5 k=10 k=20 target days
# mean F0 (st) 1.568 0.905 0.701 0.496 0.351 0.50 10
# jitter % 0.235 0.136 0.105 0.074 0.053 0.05 23
# shimmer % 1.372 0.792 0.614 0.434 0.307 0.30 21
# pause ratio 0.078 0.045 0.035 0.025 0.018 0.02 16
# final slope st/s 11.760 6.790 5.259 3.719 2.630 3.00 16十天能让平均音高精确到半个半音之内;基频微扰精确到 0.05 个百分点需要 23 天,振幅微扰精确到 0.3 个百分点需要 21 天;停顿占比精确到 0.02、句末斜率精确到 3 st/s 各需 16 天。因此,单次测量的「嗓音权威分数」除音高之外,几乎全是逐日噪声。
假说 · 尚未验证
作者的诠释(尚未验证): 至少十天录音平均得到的综合嗓音分数,比单日的同一分数更能预测听众对该说话者新录音的权威评分。什么会证明它错: 在每位说话者至少有十天录音并有独立听众评分的研究中,十天分数与评分的相关不强于单日分数。
小结
本书的指标中,只有平均音高的重复性足以从一次测量读出;微扰、停顿与斜率指标需要多天录音与近距离、固定位置的麦克风。负责任的诊断工具应公布信度,并要求重复录音。
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).