AAI_2025_Capstone_Chronicles_Combined

9

Feature Engineering

The feature engineering captures spectral and temporal properties of the harmonic complexities across each sound file. These computed features then feed into the downstream PCA model and Deep Learning model. Using Librosa and related Python audio analysis tools, first we create a custom set of forty-four descriptors for each audio sample. These include spectral centroid, bandwidth, rolloff, spectral contrast, zero-crossing rate, and Mel-frequency cepstral coefficients (MFCCs). Temporal descriptors include onset strength, attack duration, decay duration, and measures of spectral change over time. Figure 4 illustrates an example set of these descriptors for a single loaded sample.

Figure 4

A subset of engineered features showing how timbre characteristics evolve for a single sound

341