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

第三章 基频微扰与振幅微扰:这些数字测的是什么,录音设备又加进了什么

周期间的微扰

基频微扰(jitter)是声门周期逐个周期的变化,振幅微扰(shimmer)是峰值振幅逐个周期的变化;两者都是 Praat 等嗓音分析软件的标准输出 (Boersma 2001)。病理阈值常被引用为局部基频微扰 1.04%、局部振幅微扰 3.81% (Teixeira et al. 2013)。流行的建议「基频微扰低于 1%、振幅微扰低于 3% 才显得有权威」借用的正是这些临床界限。p05 为三种合成嗓音计算两项指标,并问它们有多稳定。

推导

# Chapter 3. p05: Jitter and shimmer from cycle-to-cycle periods and
# peak amplitudes. Local jitter = mean |T(i) - T(i-1)| / mean T; local
# shimmer the same for amplitudes. Three synthetic voices at 120 Hz with
# independent normal perturbations (sd of period, sd of amplitude, both
# relative): steady 0.3% and 1.5%; ordinary 0.7% and 3%; rough 1.2% and
# 5%. For independent errors local jitter is about 1.128 times the sd.
# Thresholds often quoted for pathology: jitter 1.04%, shimmer 3.81%.
# Each voice: 2,000 recordings of 50 cycles and of 500 cycles. Seed 99.
import random

def local(xs):
    d = sum(abs(a - b) for a, b in zip(xs[1:], xs)) / (len(xs) - 1)
    return d / (sum(xs) / len(xs))

VOICES = (("steady", 0.003, 0.015), ("ordinary", 0.007, 0.03),
          ("rough", 0.012, 0.05))

if __name__ == "__main__":
    rng = random.Random(99)
    print("   voice     cycles  jitter %  10-90% range    shimmer %"
          "  over a limit")
    for name, sj, sa in VOICES:
        for n in (50, 500):
            js, ss, over = [], [], 0
            for _ in range(2000):
                t = [(1 + rng.gauss(0, sj)) / 120 for _ in range(n)]
                a = [1 + rng.gauss(0, sa) for _ in range(n)]
                j, s = local(t), local(a)
                js.append(j)
                ss.append(s)
                over += j > 0.0104 or s > 0.0381
            js.sort()
            fmt = ("   {:<8}  {:6d}  {:8.2f}  {:5.2f} to {:5.2f}  {:9.2f}"
                   "  {:12.3f}")
            print(fmt.format(name, n, 100 * sum(js) / 2000, 100 * js[200],
                          100 * js[1800], 100 * sum(ss) / 2000, over / 2000))
#    voice     cycles  jitter %  10-90% range    shimmer %  over a limit
#    steady        50      0.34   0.28 to  0.40       1.68         0.000
#    steady       500      0.34   0.32 to  0.36       1.70         0.000
#    ordinary      50      0.79   0.66 to  0.92       3.39         0.170
#    ordinary     500      0.79   0.75 to  0.83       3.38         0.002
#    rough         50      1.36   1.14 to  1.58       5.63         1.000
#    rough        500      1.35   1.29 to  1.43       5.64         1.000

平稳嗓音的基频微扰是 0.34%,振幅微扰 1.68% 至 1.70%;普通嗓音 0.79% 与 3.39%;粗糙嗓音 1.36% 与 5.63%。对于独立的扰动,局部基频微扰接近周期标准差的 1.128 倍,如 p05 的注释所说。只取 50 个周期(不到半秒的发声)时,普通嗓音在 0.170 的录音中会越过两条界限之一;取 500 个周期时只有 0.002。短样本大约每六次就有一次让正常的嗓音看起来不正常。

这些指标描述的是发声的稳定性,而不是权威感。它们能区分健康与病理嗓音;但没有已发表的证据表明,在健康范围内把基频微扰从 0.79% 降到 0.34% 能使说话者更有说服力,下面的假说说明可以怎样检验。

来自录音设备的微扰

周期是在采样信号中两个事件之间测得的,所以采样网格本身就会加进微扰。p06 生成一个真实局部基频微扰为 0.34% 的嗓音,加上比信号低 40 dB 的噪声,以三种采样率采样,并在过零点测量周期。

推导

# Chapter 3. p06: Jitter that comes from the recorder. A voice at 120 Hz
# with true period jitter (sd 0.3%, local jitter about 0.34%) is drawn
# as a sine whose phase advances one cycle per glottal period, with
# noise 40 dB below the signal, and sampled at 8, 16 and 44.1 kHz.
# Periods are measured between upward zero crossings, either at the
# sample after the crossing or by linear interpolation between the two
# samples around it; a crossing within 6 ms of the last one is ignored
# (noise near zero makes extra crossings). 400 cycles per recording, 30
# recordings. Seed 99.
import random
from math import pi, sin

def record(rng, fs, n=400):
    periods = [(1 + rng.gauss(0, 0.003)) / 120 for _ in range(n)]
    edges, t = [0.0], 0.0
    for p in periods:
        t += p
        edges.append(t)
    xs, k = [], 0
    for i in range(int(t * fs)):
        ti = i / fs
        while k < n - 1 and ti >= edges[k + 1]:
            k += 1
        ph = (ti - edges[k]) / periods[k]
        xs.append(sin(2 * pi * ph) + rng.gauss(0, 0.01))
    return periods, xs

def crossings(xs, fs, interp):
    out = []
    for i in range(1, len(xs)):
        if xs[i - 1] < 0 <= xs[i] and (not out or i / fs - out[-1] > 0.006):
            if interp:
                out.append((i - 1 + xs[i - 1] / (xs[i - 1] - xs[i])) / fs)
            else:
                out.append(i / fs)
    return [b - a for a, b in zip(out, out[1:])]

def local(ts):
    return sum(abs(a - b) for a, b in zip(ts[1:], ts)) / (
        len(ts) - 1) / (sum(ts) / len(ts))

if __name__ == "__main__":
    rng = random.Random(99)
    print("   rate (Hz)  true jitter %  at sample %  interpolated %")
    for fs in (8000, 16000, 44100):
        tr, raw, lin = 0.0, 0.0, 0.0
        for _ in range(30):
            periods, xs = record(rng, fs)
            tr += local(periods) / 30
            raw += local(crossings(xs, fs, False)) / 30
            lin += local(crossings(xs, fs, True)) / 30
        print("   {:9d}  {:13.2f}  {:11.2f}  {:14.2f}".format(
            fs, 100 * tr, 100 * raw, 100 * lin))
#    rate (Hz)  true jitter %  at sample %  interpolated %
#         8000           0.34         1.00            0.42
#        16000           0.34         0.59            0.42
#        44100           0.34         0.47            0.43

若取每次过零后的那个采样点,测得的基频微扰在 8 kHz 时是 1.00%(一个其实非常平稳的嗓音,却超过了 1% 的建议值),16 kHz 时 0.59%,44.1 kHz 时 0.47%。在采样点之间插值后,三种采样率都降到 0.42% 或 0.43%:剩下超过 0.34% 的部分来自噪声,它让每次过零都略有偏移。基频微扰的数值是嗓音加上录音链的性质;只有在相同采样率、相同算法、相近噪声下比较才公平。

假说 · 尚未验证

作者的诠释(尚未验证): 在健康范围内(局部基频微扰低于 1%),一旦控制平均音高、响度与语速,基频微扰与振幅微扰对听众评价说话者的能力或可信度没有影响。什么会证明它错: 在一项预先登记的听辨实验中,把同样的语句在音高、响度与时长固定的条件下重新合成为基频微扰 0.34% 与 0.79% 两个版本,两者的能力或信任评分存在稳定差异。

小结

基频微扰与振幅微扰测的是发声的稳定性,只有在数百个周期上才可靠;临床阈值不是说服的目标。录音设备也会加进微扰:以 8 kHz 采样且不插值时,一个非常平稳的嗓音测得 1.00%。

参考文献

Boersma, Paul. 2001. “Praat, a System for Doing Phonetics by Computer.” Glot International 5 (9/10): 341 to 345.
Teixeira, João Paulo, Carla Oliveira, and Carla Lopes. 2013. “Vocal Acoustic Analysis: Jitter, Shimmer and HNR Parameters.” Procedia Technology 9: 1112 to 1122.

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).