Listeners' perceptions of the certainty and honesty of a speaker are associated with a common prosodic signature

The study investigates how listeners infer two key aspects of a speaker’s reliability—certainty (confidence in their own information) and honesty (willingness to communicate truthful information)—from speech prosody. While languages provide explicit markers of certainty, efficient cooperation also requires fast, low-level mechanisms of epistemic vigilance that generalize across languages and contexts. Prior work suggests that uncertainty and deception may share prosodic correlates associated with cognitive effort (e.g., pitch changes, slower speech, disrupted typical declination). The authors hypothesize that a common, core prosodic signature—reflecting natural signs of cognitive effort—underlies perceptions of both certainty and honesty, is perceived independently of conceptual knowledge and native language, and is extracted automatically. They further predict that although the perceptual inputs are shared, contextual and inferential processes will differentially affect judgments of dishonesty compared to certainty.

Prior elicitation studies show speakers’ uncertainty is often associated with rising intonation, decreased volume, altered mean pitch, and slower speech rate, and that some markers of deception (e.g., higher pitch, slower rate) overlap with uncertainty. However, such studies are confounded by voluntary encoding, actor stereotypes, co-variation among acoustic features, and correlational analyses that do not isolate perception mechanisms. Prosodic patterns linked to cognitive or articulatory effort (increased vocal fold tension; slower, more variable articulation; disrupted default falling pitch/volume) resemble uncertainty cues, suggesting they may be natural signs of effort rather than conventional signals. Evidence on deception cues is mixed, with speech prosody considered more reliable than other behaviors (e.g., gaze). Together, these literatures motivate testing whether a single prosodic signature of effort supports both certainty and honesty judgments and whether it generalizes across languages and beyond explicit conceptual knowledge.

Four studies (N=115) combined reverse-correlation and causal acoustic manipulations of pseudo-words. - Study 1 (N=20 French): Two-alternative forced-choice tasks one week apart. Participants heard 880 pairs of randomly prosody-manipulated bisyllabic pseudo-words (CLEESE toolbox) per session and chose which sounded more certain (task A) or more dishonest (task B), then rated confidence. Random manipulations independently modulated pitch, loudness, and duration over time within naturalistic ranges. Reverse correlation estimated normalized temporal kernels by subtracting feature trajectories of stimuli judged reliable (honest/certain) from unreliable (lying/doubtful), for 12 time points (pitch, loudness) and 5 (duration). Mixed-effects models assessed dynamic effects; psychometric curves (summed area-under-curve across dimensions) measured sensitivity; double-pass (10% repeats) and SDT modeling estimated internal noise; metacognitive sensitivity/efficiency derived from confidence-split psychometric slopes. - Study 2A (N=40 French): Causal test with context and ratings. Applied the average 3D archetypal kernels (certain, honest, doubt, lie) from Study 1 at three gains to original pseudo-words (two speakers). Separate groups rated certainty (N=20) or dishonesty (N=20) on 1–7 scales with contextual framing (confidence judgments vs poker bluff scenario). rmANOVAs tested effects of prosody type, gain, task. - Study 2B: Same participants reported conceptual beliefs about epistemic prosody (pitch, loudness, speed; static/dynamic aspects). Compared conceptual responses with perceptual ratings. - Study 3: Cross-linguistic replication of certainty ratings with same stimuli and procedure in native English (N=22) and Spanish (N=21) speakers without French exposure, plus a multi-language group (N=12; varied exposure). rmANOVA with language factor and correlations with self-reported French comprehension. - Study 4 (implicit memory; same 40 French as Study 2): 6-item auditory list of pseudo-words (1 s ISI), followed by 3-alternative forced-choice recognition of a target word (position randomized 1–6). Targets had archetypal reliable (certain/honest) or unreliable (lie/doubt) prosodies; distractors had random prosodies. Measured accuracy, RT, and confidence; mixed-effects models assessed main effects of reliability, target position, and their interaction. Outlier RTs excluded per IQR rule. Analyses pre-registered in Methods; code and data available (OSF; CLEESE).

- Common dynamic signature: Across both honesty and certainty judgments (Study 1), unreliable speech was characterized by rising intonation (significant linear segment effect for pitch; χ²=22.76, p=0.02; falling contours judged more reliable), reduced loudness at word onset (quadratic segment effect; χ²=60.18, p<0.001), and slower overall duration/speech rate (quadratic segment effect; χ²=14.55, p=0.006; global faster speech linked to reliability: certainty slope M=−0.14 SD, p=0.012; honesty M=−0.08 SD, p=0.053). Variability mattered: lower pitch variability and lower speech-rate variability predicted reliability (pitch SD difference: honesty t(18)=−4.74, p<0.001; certainty t(18)=−2.64, p=0.017; duration variability: honesty t(18)=−4.06, p<0.001; certainty t(18)=−2.94, p=0.008). Mean pitch and mean loudness were not predictive at the static level. - Strong overlap across tasks: Dynamic kernels showed no main effect of task and no segment×task interactions for any dimension; agreement across tasks on identical pairs was 61.4%±9, above chance (t(18)=5, p<0.001). Within-task double-pass agreement: certainty 69.9%±6.4 (t(18)=13, p<0.001), honesty 66.3%±9.3 (t(18)=7.4, p<0.001); honesty less consistent than certainty (t(18)=2.5, p=0.02). SDT internal noise lower for certainty than honesty (t(17)=2.23, p=0.04). Sensitivity to summed evidence above chance in both tasks (certainty Z(18)=14, p=0.0011; honesty Z(18)=33, p=0.012), but lower for honesty (Z(18)=31, p=0.01). Metacognitive sensitivity above chance in both, higher for certainty; metacognitive efficiency above chance and similar across tasks. - Causal validation with context (Study 2A): In the certainty rating task, both “certain” and “honest” archetypes were rated more certain than “doubt” and “lie” at all gains (all p<0.001; large ds). In the honesty rating task, group effects were reduced and non-linear with gain; “doubt” was often judged more dishonest than “honest” (p<0.01) and than “certain” at intermediate gain (p<0.01), but “lie” archetype was not reliably more dishonest at the group level, reflecting larger inter-individual variability and contextual inference. - Percepts vs concepts (Study 2B): Little consensus in explicit beliefs; only associations such as lower mean pitch with certainty (χ²=22.5, p<0.001) and higher mean pitch with lies (χ²=19.6, p<0.001), and higher loudness with certainty (p<0.004) were endorsed. Conceptual knowledge generally did not predict perceptual ratings; exception: concepts about speed moderated honesty ratings (F(3,39)=7.08, p=0.0007). - Cross-linguistic generalization (Study 3): French, English, and Spanish speakers showed the same certainty rating patterns (main effect of prosody F(3,180)=135.5, p<0.001, η²=0.69; no language effect F(2,60)=1, p>0.37; no interaction F(6,180)=1.4, p>0.2). Multi-language group replicated; French exposure did not correlate with effects. - Automatic impact on memory (Study 4): Words with unreliable prosody were recognized more accurately and faster, and with slightly higher confidence (accuracy and RT: reliability main effects; RT t(39)=2.31, p=0.026; confidence t(39)=−2.18, p=0.035). No type (certainty vs honesty)×reliability interaction, consistent with a common signature. Prosody effects on confidence interacted with serial position, suggesting unconscious influences when targets occurred earlier.

Findings support a core, dynamic prosodic signature—rising intonation, reduced initial loudness/accentuation, slower speech rate, and greater variability—that listeners use to judge both certainty and honesty. The signature aligns with physiological consequences of cognitive effort, indicating it constitutes a natural sign rather than a culturally learned, conventional signal. Perception of certainty maps closely onto these cues and yields precise, confident judgments. Judgments of dishonesty, however, rely on the same perceptual inputs but are less stable and more context-dependent, consistent with inferential/pragmatic accounts where listeners integrate incentives, social costs, and situational information to interpret intentions. The signature is language-independent and automatically extracted, biasing verbal working memory even when prosody is task-irrelevant, implying an adaptive vigilance mechanism in auditory processing. Decoupling between percepts and explicit concepts suggests procedural knowledge drives epistemic prosody perception, with limited conscious access to the underlying acoustic rules. These results refine models of epistemic vigilance by quantifying contributions of perception, metacognition, and inference and by distinguishing natural signs of effort from strategic signaling in social communication.

The study uncovers a language-independent, core prosodic signature of (un)reliability that supports listeners’ judgments of certainty and honesty: falling intonation, stronger initial accentuation, faster speech rate, and reduced variability signal reliability. Reverse-correlation and causal manipulations demonstrate shared perceptual representations across attitudes and languages, independence from explicit conceptual knowledge, and automatic effects on memory. Certainty judgments are tightly perceptual, whereas dishonesty judgments further depend on contextual inference and individual biases. Future work should test origins and development (infants, non-human primates), generalization to non-Indo-European and tonal languages, interactions with phrasal prosody and semantics, additional acoustic features (voice quality), and long-term memory consequences; and integrate contextual manipulations to dissociate perceptual sensitivity from decisional biases in deception detection.

- Stimuli were isolated pseudo-words without phrasal prosody or semantic content; generalization to full discourse and semantic interactions was not directly tested. - Cross-linguistic tests involved Indo-European languages; no tonal languages were included, and distant linguistic/cultural groups may reveal in-group advantages. - Sample sizes per study were modest; observed gender differences and inter-individual variations warrant larger, targeted samples to disentangle effects of gender, empathy, and anxiety. - Only pitch, loudness, and duration were manipulated; other features (e.g., voice quality) were not directly tested. - Honesty judgments showed high context-dependence; laboratory contexts may not capture real-world incentive structures. - Memory effects were short-term; impacts on long-term retention remain unknown.