Question

Does the same business-proposal text, delivered with “final falls + 0.6-second silent pauses before key high-information words”, raise blind listeners’ willingness to invest compared with a delivery of “final rises + no pauses + faster speech”? And how does it affect perceived persuasiveness and credibility?

Stimuli

A fictional business proposal of about 90 seconds (the funding pitch of a fictional company) is recorded in a natural reading by each of four voice actors (two women, two men). An open-source voice-transformation tool such as CLEESE (Burred et al. 2019) is then used to make two versions of each actor’s recording:

• Version A: every statement ends with a pitch fall of at least 2 semitones; a 0.6-second silent pause is inserted before each of 6 pre-marked high-information words (figures, key conclusions); rate unchanged.
• Version B: every statement ends with a pitch rise of at least 2 semitones; no pauses inserted; rate increased by 15 %.

The two versions differ in total length (version B is shorter); this is a property of the combination itself. Before the experiment, a research assistant who is not involved in recruiting listeners checks by acoustic measurement that every stimulus meets these conditions (the manipulation check).

Design

Sample size

# Sample size for the double-blind randomised trial. Listeners hear ONE version of the same
# 90-second business pitch, read by one of four voice actors (2 women, 2 men) and processed so that
# only prosody differs:
#   A: falling final contours + a 0.6-s silent pause before each high-information word
#   B: rising final contours + no added pauses + 15 % faster speech
# Primary outcome: willingness to invest, the share of a hypothetical 10,000-unit budget allocated
# to the venture (0-100 %). Effect in SD units; normal approximation; two-sided alpha 0.05.
import math
from statistics import NormalDist
N = NormalDist()

def n_per_arm(d, alpha=0.05, power=0.90):
    return math.ceil(2 * (N.inv_cdf(1 - alpha / 2) + N.inv_cdf(power)) ** 2 / d ** 2) + 1

for d in (0.20, 0.25, 0.30, 0.40):
    print(f"standardised effect {d:.2f}: {n_per_arm(d, power=0.8)} per arm for 80 % power, {n_per_arm(d)} for 90 %")
n = n_per_arm(0.25)
# Randomisation is stratified by voice (each voice reads both versions, equally often in each arm),
# and voice enters the analysis as a fixed blocking factor, so it costs no extra sample. The price:
# with only 4 voices the result generalises to other voices only weakly.
print(f"chosen: 0.25 SD, 90 % power -> {n} per arm ({n // 4 + 1} per voice and arm); recruit"
      f" {math.ceil(n / 0.85)} per arm allowing 15 % exclusions, {2 * math.ceil(n / 0.85)} in all")
# standardised effect 0.20: 394 per arm for 80 % power, 527 for 90 %
# standardised effect 0.25: 253 per arm for 80 % power, 338 for 90 %
# standardised effect 0.30: 176 per arm for 80 % power, 235 for 90 %
# standardised effect 0.40: 100 per arm for 80 % power, 133 for 90 %
# chosen: 0.25 SD, 90 % power -> 338 per arm (85 per voice and arm); recruit 398 per arm allowing 15 % exclusions, 796 in all

With a smallest effect of interest of 0.25 standard deviations and 90 % power, each arm needs 338 analysable participants (about 85 per voice actor and version); allowing for about 15 % to be excluded for failing attention checks, each arm recruits 398, 796 in all. Because randomisation is stratified by voice actor and voice actor enters the analysis as a fixed factor, the actors themselves cost no extra sample; the price is that with only four actors the results generalise only weakly to other voices.

Decision rule

# Decision rule, fixed before data collection. d = version A (falling + pauses) minus version B
# (rising + no pauses + faster) in willingness to invest, standardised, with its 95 % CI.
SESOI = 0.25

def verdict(d, lo, hi, manip_ok):
    if not manip_ok:
        return f"effect {d:+.2f}: NOT INTERPRETABLE - the stimuli failed the acoustic manipulation check"
    if lo >= SESOI:   v = "version A raises willingness to invest by at least the smallest effect of interest"
    elif lo > 0:      v = "version A raises it; size not shown to reach 0.25 SD"
    elif hi < 0:      v = "version A LOWERS willingness to invest"
    elif lo >= -SESOI and hi <= SESOI: v = "no difference worth acting on"
    else:             v = "unclear"
    return f"effect {d:+.2f} ({lo:+.2f} to {hi:+.2f}): {v}"

# Manipulation check on the stimuli themselves (measured before listeners are recruited): every
# version-A sentence must end with a fall of at least 2 semitones and every version-B declarative
# with a rise of at least 2 semitones; pauses must be 0.6 +/- 0.05 s.
for r in ((0.41, 0.27, 0.55, True), (0.12, -0.03, 0.27, True), (0.05, -0.10, 0.20, True), (0.30, 0.15, 0.45, False)):
    print(verdict(*r))
print("Credibility and persuasiveness ratings are secondary: reported with intervals, never decisive.")
# effect +0.41 (+0.27 to +0.55): version A raises willingness to invest by at least the smallest effect of interest
# effect +0.12 (-0.03 to +0.27): unclear
# effect +0.05 (-0.10 to +0.20): no difference worth acting on
# effect +0.30: NOT INTERPRETABLE - the stimuli failed the acoustic manipulation check
# Credibility and persuasiveness ratings are secondary: reported with intervals, never decisive.

The example results are invented. If the stimuli fail the acoustic manipulation check, the report says “not interpretable” whatever the results, and the trial is rerun with corrected stimuli. Only when the lower limit of the confidence interval reaches 0.25 standard deviations will the report say that “version A raised willingness to invest by at least the smallest effect of interest”.

What this trial cannot show

First, versions A and B change three features at once (final movement, pauses, rate), in order to test a “package” of prosodic advice; even a positive result cannot show which feature did the work, and a 2×2×2 factorial design would be needed to separate them. Second, willingness to invest is an allocation of a hypothetical budget, not a real investment, and participants know they are in a study. Third, with four voice actors, one text and one language (English and Chinese versions must be run separately), the results generalise only weakly to other voices, texts and languages. Fourth, research discussed in Chapters 1 and 3 suggests that a rise may be judged more trustworthy in some contexts, so version B is not necessarily the “worse” version; the trial may find no difference in credibility, or even higher credibility for version B, and the decision rule reports such results just as faithfully.

Ethics and conflicts of interest

The trial requires approval from an independent ethics committee and informed consent, and the protocol and analysis plan are to be registered before the first participant is enrolled (Nosek et al. 2018). After the study, participants are told its true purpose (that the prosody of the voice had been altered) and may ask for their data to be withdrawn. Voice actors agree in advance that their recordings may be digitally processed and used for research.