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

第四章 句末降调与战略停顿:两个韵律信号的测量

陈述句的句末降调

英语和普通话的陈述句通常以音高下降结束;句末上扬(uptalk)常被听成不确定。句末降调的斜率说起来容易、测起来难,因为音高追踪会出现倍频错误:在嘎裂或气息声的结尾,追踪器可能报出真实频率的两倍或一半。p07 用两种估计方法测量合成的降调、平调与升调结尾的最后 300 毫秒。

推导

# Chapter 4. p07: Measuring the final fall of a statement. The last 300
# ms of an utterance in 10 ms frames, pitch in semitones: a fall of 6
# semitones (-20 st/s), a level ending, or a rise of 4 (uptalk). The
# pitch tracker adds noise of sd 0.5 st per frame and, with chance q
# per frame, an octave error of +12 or -12 st. The slope is estimated
# by least squares (OLS) or by the median of pairwise slopes
# (Theil-Sen) and called "fall" below -5 st/s, "rise" above +5, else
# "level". 2,000 utterances per case. Seed 99.
import random

T = [i * 0.01 for i in range(30)]
SHAPES = (("fall", -20.0), ("level", 0.0), ("rise", 4 / 0.29))

def ols(ys):
    mt, my = sum(T) / 30, sum(ys) / 30
    return sum((t - mt) * (y - my) for t, y in zip(T, ys)) / sum(
        (t - mt) ** 2 for t in T)

def theil_sen(ys):
    s = sorted((ys[j] - ys[i]) / (T[j] - T[i])
               for i in range(30) for j in range(i + 1, 30))
    return s[len(s) // 2]

def label(b):
    return "fall" if b < -5 else "rise" if b > 5 else "level"

if __name__ == "__main__":
    rng = random.Random(99)
    print("   octave errors  OLS wrong  Theil-Sen wrong  (share of calls)")
    for q in (0.0, 0.02, 0.05, 0.10):
        bad = [0, 0]
        for _ in range(2000):
            name, slope = SHAPES[rng.randrange(3)]
            ys = []
            for t in T:
                y = slope * t + rng.gauss(0, 0.5)
                if rng.random() < q:
                    y += rng.choice((12, -12))
                ys.append(y)
            bad[0] += label(ols(ys)) != name
            bad[1] += label(theil_sen(ys)) != name
        print("   {:13.2f}  {:9.3f}  {:15.3f}".format(
            q, bad[0] / 2000, bad[1] / 2000))
#    octave errors  OLS wrong  Theil-Sen wrong  (share of calls)
#             0.00      0.000            0.000
#             0.02      0.076            0.000
#             0.05      0.152            0.000
#             0.10      0.227            0.002

没有倍频错误时,两种方法都能正确判断每个结尾。当 0.02 的帧出现倍频错误时,普通最小二乘法会误判 0.076 的结尾;0.05 时误判 0.152;0.10 时误判 0.227。成对斜率的中位数(Theil-Sen 估计)几乎不受影响:0.10 时只误判 0.002。任何按句末降调陡峭程度给说话者打分的规则,都应使用稳健的斜率,否则打分测到的主要是追踪器。

什么算停顿

常有人说,有说服力的说话者「战略停顿」占说话时间的 18% 至 25%。p08 说明,这样的数字离开定义就无法解读:多低的声级算静音,静音要多长才算停顿。

推导

# Chapter 4. p08: How a "pause ratio" depends on its definition. Sixty
# seconds of synthetic speech in 10 ms frames: phrases of 1.5 to 4 s
# separated by pauses (log-normal, median 450 ms); inside phrases,
# syllables at 5 a second, each followed with chance 0.3 by a stop
# closure of 40 to 120 ms. Speech frames have levels around 65 dB (sd
# 6); silent frames sit at a 35 dB noise floor (sd 3). A frame is
# silent if its level is below a threshold; the pause ratio is the
# share of time in silent runs at least as long as a minimum. 200
# recordings, seed 99.
import random
from math import exp

def recording(rng):
    frames = []
    while len(frames) < 6000:
        end = len(frames) + rng.randint(150, 400)
        while len(frames) < end:
            frames += [True] * rng.randint(12, 28)
            if rng.random() < 0.3:
                frames += [False] * rng.randint(4, 12)
        frames += [False] * max(1, round(45 * exp(rng.gauss(0, 0.5))))
    return [rng.gauss(65, 6) if s else rng.gauss(35, 3)
            for s in frames[:6000]]

def ratio(levels, thr, min_ms):
    total, run = 0, 0
    for v in levels + [99.0]:
        if v < thr:
            run += 1
        else:
            if run * 10 >= min_ms:
                total += run
            run = 0
    return total / len(levels)

if __name__ == "__main__":
    rng = random.Random(99)
    recs = [recording(rng) for _ in range(200)]
    mins = (0, 100, 150, 250, 400)
    print("   threshold (dB)  minimum pause (ms):")
    print("   " + " " * 16 + "".join("  {:>8d}".format(m) for m in mins))
    for thr in (45, 50, 55):
        row = [sum(ratio(r, thr, m) for r in recs) / 200 for m in mins]
        print("   {:14d}  ".format(thr) + "".join(
            "  {:8.3f}".format(x) for x in row))
#    threshold (dB)  minimum pause (ms):
#                             0       100       150       250       400
#                45       0.241     0.195     0.157     0.152     0.127
#                50       0.245     0.196     0.157     0.153     0.128
#                55       0.277     0.197     0.157     0.153     0.129

同一批录音,若每个静音帧都算,停顿占比是 0.241;若排除短于 100 毫秒的静音(多半是词内塞音的闭塞段),是 0.195;150 毫秒时 0.157,250 毫秒时 0.152,400 毫秒时 0.127。静音阈值在这里影响较小(不设最短时长时,50 dB 为 0.245,55 dB 为 0.277),但这只是因为合成的噪声底远低于语音。「停顿占比 18% 至 25%」可以按不同的定义把同一个说话者描述为 0.127 或 0.241;一个区间只有附上最短停顿时长才有意义。

假说 · 尚未验证

作者的诠释(尚未验证): 若把停顿定义为至少 250 毫秒的静音,在公开演讲中被评为最有说服力的说话者,其停顿占比落在宽度不超过 0.10 的区间内,而该区间的中心因语言而异。什么会证明它错: 在附有评分的多语言演讲语料中,评分最高的十分之一说话者的停顿占比跨度超过 0.10,或其中位数在不同语言间没有差异。

小结

一旦音高追踪出现倍频错误,测量句末降调就需要稳健的斜率估计。停顿占比主要取决于最短停顿时长:同一批录音得到 0.241 或 0.127,所以没有定义的目标区间不是测量。

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).