Amaç Erdem & Musical Intelligence

Musical Intelligence (MI) connects acoustic structure, temporal context and cognitive-state updating in one inspectable framework. It takes raw audio at 44.1 kHz and produces evolving model states with anatomical and neurochemical annotations. These mappings express literature-derived hypotheses; their values are computational outputs.

Published, adapted and author-specified formulations are documented together. They include Sethares roughness, precision-weighted belief updates and hypotheses about anticipatory and consummatory reward. The fixed feature engine makes each transformation traceable; some behavioural and neural evaluations use fitted statistical readouts to relate those features to observations.

The source package and evaluation archive document equations, parameter inventories, protocols and saved results. Each comparison retains its target, sampling unit and analysis choices. The revised account distinguishes implementation checks, recomputed observations and experiments that still require independent evaluation.

Three composable layers connect sound to temporal features and cognitive hypotheses. Their inputs, intermediate states and output links can be traced through a common execution graph.

• R³ — RECEPTIVE RESONANCE REPRESENTATION

  97 perceptual outputs across nine groups, including consonance, energy, timbre, rhythm and harmony. Fixed signal-processing operations combine published, adapted and author-specified formulations, providing an acoustic reference for evaluating the later transformations.

• T³ — TEMPORAL TENSOR TOPOLOGY

  Temporal operators integrate features across 32 horizons, from milliseconds to minutes. Past, forward and centred windows make temporal context explicit; forward and centred windows use future samples. The cognitive layer requests the feature combinations it needs.

• C³ — COMPUTATIONAL COGNITIVE CORE

  89 mechanisms update 131 Bayesian beliefs across eight functional domains. Declared connections map their outputs to 26 anatomically labelled regions and dopamine, norepinephrine, opioid and serotonin proxies. These pathways specify computational hypotheses about cognition, anatomy and neuromodulation.

The evaluation archive relates selected model outputs to behavioural ratings and public fMRI recordings, alongside implementation and literature-consistency checks. The revised manuscript reports the positive associations and the conditions under which they weaken, giving each result its own target and scope.

• Selected soundtrack outputs reach ρ = 0.74 for sadness across 110 rating rows representing 102 unique pieces. Continuous tension tracking averages ρ = 0.42 across 38 pieces. These associations depend on channel selection and, for tension, a per-piece lag search; prediction on new music remains a subsequent test.

• Regional fMRI comparisons report 16 / 22 registered target pairs passing FDR control in 17 participants, with 21 pairs evaluable. Independent stimulus timing is still needed. A separate four-participant study uses within-subject clip cross-validation; selection of voxels by their held-out scores limits its performance estimates.

• Regional output magnitudes are stable across two audio collections (r = 0.998). This describes the model's internal organisation. Regional model–BOLD performance correlates at r = 0.237 (p = 0.149), leaving empirical transfer across datasets unestablished.

• Eleven literature-consistency checks examine the neurochemical pathways. The archived caudate-labelled signal leads the NAcc-labelled signal on 52 / 56 tracks across ChillsDB, DEAM and Eerola. These are model dynamics; physiological and drug-response measurements remain external tests of the proposed correspondences.

• Chill-event associations vary with the analysis pipeline, PMEmo emotion tracking is weak, and the cross-cultural composite fails its criterion. Audio-native Cheung reward estimates await chord-onset reconciliation. These results identify specific targets for refinement and independent testing.

MI provides a common computational setting for comparing perceptual, temporal and cognitive accounts of music. Researchers can inspect temporal profiles, select excerpts that differ in a candidate feature, and specify measurable predictions about listener responses.

The next question is which intermediate transformations improve prediction beyond acoustic features on new music and listeners. Matched baselines, component ablations and independently aligned observations can test that contribution. Individual differences in familiarity, preference and training remain an important extension of the current model.