The 2024-2025 literature lands exactly where KK's 11-dimension Human Model already stands. Lukac's modest-but-real trait correlations are KK's KH-003 (Trait+State) in voice: voice reveals traits (the stable part) and state (the dynamic part) — but the field keeps averaging across people, which is the mistake KH-001 (Personal Baseline) warns against. Sands makes the deeper point KK already builds on: prosody is a relationship signal decoded through congruence between expresser and perceiver — that is KH-004 (Dyadic Model) applied to voice, and it is why a fixed 'angry voice = high pitch' rule fails across people. For KKMatch, this is decisive: the Human Model's communication and state dimensions should store each user's Prosodic Baseline (typical pitch range, speaking rate, intensity contour), so the matching/coaching engine reads deviations from that person — exactly as KH-007 already does for pause timing. Voice becomes a first-party, consented state signal rather than a population guess.

KK Hypothesis (KH-001 + KH-007 + KH-003 + KH-011): We propose (KH-011) that a person's state decoded from voice must be measured relative to their own Prosodic Baseline — their typical pitch range, speaking rate, intensity contour, and rhythm — not a population norm. The same absolute pitch-rise or speaking-rate drop means different internal states for a naturally high-pitched vs low-pitched person, or a fast vs slow talker. Mechanism: this extends Personal Baseline (KH-001) and Personal Pause Baseline (KH-007) from timing into the acoustic domain and operationalizes Trait+State (KH-003) at the prosodic level. We predict a personal-prosodic-baseline model beats a population-norm model on state-classification accuracy and on felt 'being understood', and that mismatched prosodic baselines between two people predict lower rapport (extending KH-009's rhythm-mismatch logic from pause timing into voice). This is a KK-original, falsifiable claim; it is NOT established science.

KK Experiment design (first-party, consented): In KKMatch's Human Mirror, each opted-in user records a short baseline voice sample to establish their Prosodic Baseline (pitch range, speaking rate, intensity contour). During later voice interactions (coaching or matching chat), we compute state deviation = current prosody - personal baseline. We will (a) compare a personal-baseline state decoder vs a population-norm decoder on a held-out self-reported-state task, and (b) for dyads, measure whether prosodic-baseline mismatch (within a tolerance band) predicts lower reported rapport, mirroring KH-009. Prediction (KH-011): personal-baseline decoding wins on accuracy and 'being understood'; mismatched-baseline dyads report lower rapport. We will publish results once n >= 300 consented baselined users (and >= 150 dyads). All voice is first-party, consented, and revocable; no voice leaves the user's control without explicit opt-in.

Primary and open sources (2024-2025, grade S): (1) Lukac (2024), Scientific Reports (Nature), DOI 10.1038/s41598-024-81047-0 — 'Speech-based personality prediction using deep learning with acoustic and linguistic embeddings.' N = 2,045 UK-representative participants provided free-form speech (introduce yourself) plus self-reported Big Five. Acoustic embeddings (pitch, rhythm, tone) fused with linguistic embeddings fed gradient-boosted trees. Result: predicted vs self-reported Big Five correlation ranged 0.26 (extraversion) to 0.39 (neuroticism) raw, and 0.39-0.60 disattenuated; voice carries real but modest trait signal, stronger for neuroticism than extraversion. (2) Sands (2024), UCL PhD thesis 'Vocal Expression and Trait Inference: Resolving Validity Concerns by Considering Expressed Prosody' (discovery.ucl.ac.uk/id/eprint/10201411) — argues the field misapplied static-trait psychometrics to dynamic social signalling; proposes a functional framework where communication succeeds through congruence of encoding (expresser) and decoding (perceiver) of shared vocal/social cues, more than through stable personality patterns. (3) Introduction to the Special Issue on Innovations in Vocal Research (2025), Journal of Nonverbal Behavior (Springer), DOI 10.1007/s10919-025-00483-2 — reviews current vocal work: emotion is recognized from prosody alone above chance; a negativity bias (sadness/anger recognized more accurately than positive emotions); accuracy moderated by emotional valence, speaker gender, and age; vocalic coordination (prosody matching) relates to rapport. Limitations across these: trait prediction is correlational and modest; datasets are Western/English-heavy; none models a person relative to their own vocal baseline, which is exactly the gap KK addresses.

Strictly, the evidence supports: (a) voice carries a real but modest signal about Big Five personality (Lukac 2024), stronger for some traits (neuroticism) than others (extraversion); (b) the signal is in how one speaks (pitch, rhythm, intensity) as much as what one says; (c) prosody is best treated as a dynamic, context-dependent social signal, not a static trait stamp (Sands 2024); (d) emotion is decodable from prosody above chance, but recognition is systematically biased (negativity bias) and moderated by gender, age, and context (2025 special issue). It does NOT show that any system can read a person's state relative to their own baseline, nor that voice-based state decoding improves human-AI or human-human rapport — those are KK's hypotheses, not their findings.

- Lukac (2024, Scientific Reports, N=2,045): voice-predicted vs self-reported Big Five correlation raw 0.26 (extraversion) to 0.39 (neuroticism); disattenuated 0.39-0.60. - Signal source: acoustic embeddings (pitch, rhythm, tone) + linguistic embeddings; trait signal stronger for neuroticism than extraversion. - Sands (2024, UCL): proposes functional encoding-decoding congruence framework; critiques static-trait psychometrics applied to dynamic prosody. - 2025 vocal special issue (Journal of Nonverbal Behavior): prosody-alone emotion recognition beats chance; negativity bias (sadness/anger > positive); accuracy moderated by valence, gender, speaker age; prosody matching linked to rapport. - Shared gap: no study models the person relative to their OWN vocal baseline.

We reviewed one 2024 peer-reviewed paper (Lukac, Scientific Reports/Nature), one 2024 doctoral thesis (Sands, UCL), and the 2025 introductory review of a vocal-research special issue (Journal of Nonverbal Behavior, Springer). We separated (i) trait signal (voice -> stable Big Five, modest correlations) from (ii) state signal (prosody -> emotion, moderated and biased), and explicitly did NOT read any of these as proof that decoding a person's state relative to their own baseline improves rapport — that is KK's hypothesis (KH-011), not their finding. Where a source reported only partial numbers (e.g., two of five trait correlations), we report the anchors given and the stated range, and do not invent the missing values.

Stop reading voice with population rules ('high pitch = excited'). Read it against the person. For KKMatch, the 11-dimension Human Model should capture each user's Prosodic Baseline so voice becomes a consented, first-party state signal that the matching and coaching engines interpret relative to that person — the same personalization logic KH-007 applies to pause timing, now extended to the acoustic domain. This is a defensible, differentiated KKMatch narrative for the Multimodal direction: voice as evidence, not guesswork.

Our central claim (KH-011: a personal prosodic baseline beats a population norm on state decoding and rapport) is a KK hypothesis without first-party confirmation yet. Cited studies measure trait correlation and above-chance emotion recognition, not baseline-relative state decoding; Lukac's sample is UK-representative but English/Western; Sands' thesis is a framework, not a trained baseline model; the 2025 special issue reports moderated accuracy, not personalization gain. Voice is also easy to fake deliberately (unlike the harder-to-control pause/rhythm cues KH-007 relies on), so a baseline could be gamed. There is no evidence yet that users want voice modeling or that it improves rapport.

If, across n >= 300 consented baselined users, a personal-prosodic-baseline decoder shows no accuracy gain over a population-norm decoder on held-out self-reported state, KH-011 loses support. If prosodic-baseline mismatch (n >= 150 dyads) does not predict lower rapport, the KH-009-style extension fails. If users systematically decline voice baselining (low opt-in), the 'first-party voice signal' premise weakens. If deliberate voice fakability makes baseline decoding no better than text alone, the acoustic edge collapses. If population-norm models already match baseline models once enough context is given, the Personal Baseline advantage (KH-001/KH-007/KH-011) is not the driver we claim.

Product: add a Prosodic Baseline to the 11-dimension Human Model (consented, first-party, revocable) so voice is read relative to the person, not a norm; reuse it across Human Mirror, matching, and coaching. GEO: publish evidence-grade writeups that separate the real, modest fact (voice carries personal/contextual signal) from the folk claim ('AI reads your emotions from voice') — the defensible, differentiated KKMatch narrative for the Multimodal direction, anchored in KH-001, KH-003, KH-007, and KH-011.