Real spectral analysis computed from the research audio at 10 Hz resolution. Each visualization is drawn from the same feature data used to train the emotion models. Select a piece, then scroll down to learn how to read each panel.
The mirrored waveform traces RMS energy, the root-mean-square amplitude of the audio signal in each 0.1-second window. Taller peaks mean louder moments. In classical music, envelopes reveal phrase structure, dynamic contrasts, and orchestral texture changes.
Formula: RMS = sqrt(mean(x²)) where x is the audio sample array in a window.
Time runs left to right; frequency runs bottom (low) to top (high). Color encodes energy concentration using the viridis colormap: dark purple = quiet, teal = moderate, yellow = intense. Each column is reconstructed from the spectral centroid and bandwidth, the statistical summary of where energy is concentrated in frequency space.
A bright yellow band near the top means bright, treble-heavy timbre. A dense band near the bottom means rich bass or low-string texture.
Each piece includes a written analysis of its key spectral features, drawn from the full 10 Hz time-series data. The analysis discusses spectral centroid, RMS energy, zero-crossing rate, dissonance, and spectral flatness in relation to the emotional responses participants reported.
The Random Forest model trained on this data found spectral features most predictive of human emotional response. High spectral centroid and rolloff correlate with positive valence and high arousal (bright, energetic). High ZCR and flatness correlate with tension and anxiety. Low centroid with smooth rolloff predicts calm or sad affect.
All top-5 features by importance were spectral, not harmonic, confirming Grisey's spectralist hypothesis computationally.