Instruction information and spectral sensitivity in a musical model of speech and song

A model can retain information about vocal expression without assigning its native affect scores the expected direction. We examined this distinction in Musical Intelligence (MI), a system designed for musical listening, using an instruction factorial and a subsequent acoustic intervention. The first study analysed 342 synthetic recordings from 172 requests. Across held-out sentences, a fixed probe decoded five speech instructions from 889 C³ dimensions with 68.9% balanced accuracy, compared with 57.8% from 15 acoustic descriptors and 35.6% from two native affect outputs. Yet native valence was lower for excited than sad speech in every text. A second study selected six recordings and independently varied an instrument-derived upper-harmonic envelope and residual retention, yielding 1,596 primary full-recording evaluations. Mean native valence rose from 0.139 to 0.435 for cello-derived and 0.437 for viola-derived envelopes, while all 588 primary paired-anchor contrasts retained the inverse instruction ordering. Residual removal produced a substantially larger shift than envelope transfer alone. Restoring the source consonance-feature trajectories suppressed 94.8% and 94.2% of the mean endpoint rises. Arousal ordering depended on waveform level, and 192 near-endpoint probes exposed substantial score differences between very close waveforms. These results distinguish decodable instruction information, native affect mapping and numerical sensitivity. They identify a computational route through which speech-derived texture changes affect scores, without establishing consonant-specific causality, natural instrument equivalence or human perceptual thresholds.