AAI_2025_Capstone_Chronicles_Combined

21

Unexpected Findings

One unexpected finding is that the PCA model performed better than anticipated in several retrieval scenarios, especially for samples with pronounced spectral differences or clear envelope behavior. This suggests that traditional MIR descriptors remain useful for representing broad timbral categories and may serve as a computationally efficient alternative for large-scale indexing. Another unexpected finding was that retrieval quality remained strong even when validation metrics suggested mild overfitting. Although the model did not fully optimize its loss, the system still surfaced perceptually coherent neighbors, suggesting that timbre similarity emerges in the embedding even without perfect label prediction.

Implications for Real-World Use

The system ultimately aims to support creative audio workflows. The results show that learned timbre embeddings provide a flexible and intuitive search mechanism, while PCA provides a transparent baseline that helps interpret timbral dimensions. Together, these models contribute to a system that can reduce time spent searching through large audio libraries and help artists explore timbre in an intuitive way.

Future Work

One extension of the work involves expanding the dataset to include more varied audio sources. The NSynth dataset provides a controlled soundset with consistent pitched recordings, although it does not represent speech, polyphonic sound mixtures, environmental or natural

353