5D Library › Acoustic Persuasion, Prosodic Information Theory and Interpersonal Physiological Synchrony · 中文
Chapter 1 Paralinguistic Acoustics: Why “How” Rewrites the Credibility of “What”
A misused number
Almost every bestselling book on public speaking mentions a number: of the effect of communication, words account for 7 %, tone of voice for 38 % and facial expression and body language for 55 %. It is usually attributed to the psychologist Albert Mehrabian.
Evidence
Where the number comes from. Mehrabian and Wiener had participants hear single words (words conveying liking, neutrality or dislike) spoken in different tones, and judge the speaker’s attitude towards the listener (Mehrabian and Wiener 1967); Mehrabian and Ferris combined recordings of the single word “maybe” spoken in different tones with photographs of faces showing different attitudes, again asking participants to judge attitude (Mehrabian and Ferris 1967). From these two studies Mehrabian derived a combination formula: when the words, tone and facial expression conveying a liking or disliking are inconsistent, participants’ judgements of the speaker’s liking combine the three channels with weights of roughly 0.07, 0.38 and 0.55 (Mehrabian 1971).
The formula therefore has three strict limits: it concerns only liking or disliking, not facts, arguments or any complex information; it applies only when the channels contradict each other; and its “words” are single words, not sentences or passages. Extending it to “words are only 7 % of communication” is a serious misreading of the original research, as Mehrabian himself has repeatedly pointed out.
But behind the misreading lies a real phenomenon: the same words spoken in different voices do lead listeners to judge the speaker differently. This chapter discusses which features of the voice have this effect, and how large it is.
What the voice reveals
Evidence
Hearing versus reading. Schroeder and Epley gave evaluators the same self-presentation by job candidates in different forms: a written transcript, an audio recording or a video. Evaluators who heard the voice rated the candidates’ competence, thoughtfulness and their own interest in hiring them more highly than evaluators who only read the text, even though the words were identical (Schroeder and Epley 2015). They argued that the variation in a speaking voice conveys cues that the speaker is “thinking”.
The most studied acoustic features are:
- Fundamental frequency (): how many times a second the vocal folds vibrate, heard as pitch.
- Formant dispersion: the mean spacing between adjacent formants. Formants are set by the shape and length of the vocal tract; the longer the tract, the closer together the formants. Formant dispersion correlates with vocal-tract length and body size (Fitch 1997), and listeners use it as a cue to size and dominance (Puts et al. 2007).
- Harmonics-to-noise ratio (HNR): the ratio of periodic harmonic energy to aperiodic noise in the voice, roughly corresponding to “clear” versus “hoarse or breathy”.
- Speech rate: syllables per second, and the length and placement of pauses.
Evidence
Pitch and judgements of leadership. Klofstad and colleagues raised and lowered the pitch of recordings of the same campaign slogan and asked participants which version they would rather vote for. Whether the speaker was a man or a woman, participants tended to choose the lower-pitched version; lower-pitched voices were also rated stronger and more competent (for male voices, by male listeners only) (Klofstad et al. 2012). Puts and colleagues found that both lower pitch and smaller formant dispersion made men’s voices sound more dominant (Puts et al. 2007).
Evidence
Rate and pitch. Apple and colleagues varied speech rate and pitch independently by speech synthesis and found that slower speakers were judged less truthful, less persuasive and more passive, and higher-pitched speakers less truthful, less emphatic, less potent (weaker) and more nervous (Apple et al. 1979). This is a reminder that “the slower the better” is false: both too slow and too fast have costs.
Evidence
The “social code” of prosody. Ponsot and colleagues used reverse correlation: participants judged the trustworthiness and dominance of many recordings of bonjour (“hello”) whose pitch contour had been randomly varied, from which the listeners’ mental “prototypes” were reconstructed. The prototype judged dominant had lower pitch and a falling end; the prototype judged trustworthy had pitch that rose at the end of the word (Ponsot et al. 2018). The same prosodic feature can thus produce opposite judgements on different dimensions; the finding comes from a single French word, and whether it carries over to the ends of statements is untested.
Estimating “vocal-tract length” from formants
Calculation
Approximate the vocal tract as a uniform tube closed at one end (the glottis) and open at the other (the lips): formants then lie at , and adjacent formants are apart. From measured formants an “apparent vocal-tract length” can be inferred. The program also converts common pitch changes into semitones: .
# Two acoustic cues listeners use to judge a speaker's size and authority.
# (1) Fundamental frequency f0: changes are perceived roughly logarithmically, so they are expressed
# in semitones: 12 * log2(f_new / f_old).
# (2) Formant dispersion Df (Fitch 1997): the mean spacing between successive formants. For a
# uniform tube closed at the glottis and open at the lips, formants sit at (2n - 1) c / (4L), so
# their spacing is c / (2L) and an apparent vocal-tract length is L = c / (2 Df).
# The formant values below are illustrative, for a neutral vowel.
C = 35000.0 # speed of sound in warm moist air, cm/s (about 350 m/s)
import math
def semitones(f_new, f_old):
return 12 * math.log2(f_new / f_old)
def dispersion(formants):
return sum(formants[i + 1] - formants[i] for i in range(len(formants) - 1)) / (len(formants) - 1)
voices = {"speaker A": [520, 1540, 2560, 3570], "speaker B": [570, 1700, 2830, 3960],
"speaker C": [640, 1900, 3160, 4420]}
print("voice formants F1-F4 (Hz) dispersion Df apparent vocal-tract length")
for name, F in voices.items():
df = dispersion(F)
print(f"{name:11s} {str(F):26s} {df:9.0f} Hz {C / (2 * df):9.1f} cm")
print("check: uniform 17.5 cm tube ->", [round((2 * n - 1) * C / (4 * 17.5)) for n in range(1, 5)], "Hz")
# How large are the pitch changes used in perception experiments?
for f_old, f_new in ((120, 110), (120, 100), (220, 200), (220, 185)):
print(f"f0 {f_old} -> {f_new} Hz: {semitones(f_new, f_old):+.2f} semitones ({(f_new / f_old - 1):+.0%})")
# voice formants F1-F4 (Hz) dispersion Df apparent vocal-tract length
# speaker A [520, 1540, 2560, 3570] 1017 Hz 17.2 cm
# speaker B [570, 1700, 2830, 3960] 1130 Hz 15.5 cm
# speaker C [640, 1900, 3160, 4420] 1260 Hz 13.9 cm
# check: uniform 17.5 cm tube -> [500, 1500, 2500, 3500] Hz
# f0 120 -> 110 Hz: -1.51 semitones (-8%)
# f0 120 -> 100 Hz: -3.16 semitones (-17%)
# f0 220 -> 200 Hz: -1.65 semitones (-9%)
# f0 220 -> 185 Hz: -3.00 semitones (-16%)The program’s three illustrative voices have formant dispersions from about 1,020 Hz to 1,260 Hz, corresponding to apparent vocal-tract lengths from about 17 cm to 14 cm. Listeners cannot “hear” tract length directly, but they are sensitive to the overall position of the formants, and associate lower, closer-spaced formants with larger size and greater dominance (Puts et al. 2007). The second half of the program converts pitch changes of 10–35 Hz into semitones: about 1.5–3 semitones (8 %–17 %), smaller than many people imagine.
Note
How large these effects are, and for whom. Most of the effects above come from laboratory comparisons of digitally processed recordings of the same speaker; they are average effects, individuals differ greatly, and most of the research was done in Western countries. Lower pitch is not simply better: deliberately pushing the voice down, away from one’s comfortable range, can strain the voice and can sound unnatural. This book does not recommend that anyone deliberately change their pitch.
Hypothesis · not established
The channel-consistency hypothesis (not yet established). The effect of the voice on a speaker’s credibility is smaller when content and vocal cues agree and larger when they disagree (exactly the special condition of Mehrabian’s experiments). A testable implication: in the same business statement, the fall in listeners’ credibility ratings when the text states a firm fact while the prosody expresses hesitation should be larger than the fall caused by prosody changing from “better” to “worse” when text and prosody both express certainty. If the two falls are similar, the hypothesis is weakened.
Summary
“7 %–38 %–55 %” applies only to the experimental situation in which a single word conveys liking and the channels contradict each other; it is not a general law of communication. But the voice does affect listeners’ judgements: hearing a voice rather than reading the same words leads evaluators to rate competence more highly; lower pitch and closer formants go with judgements of dominance; speaking too fast and too slow both carry costs; and the same prosodic feature can produce opposite judgements on different dimensions. The next chapter takes up a more specific question: why do pauses have power?
References
Fan, Y. (2026). Acoustic Persuasion, Prosodic Information Theory and Interpersonal Physiological Synchrony: Pitch Contours, Pause Entropy and Listener Autonomic Coupling in Public Speaking and Leadership [声音说服力、韵律信息论与人际生理同步] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.23753267 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).