AAI_2025_Capstone_Chronicles_Combined

13

These vectors place each audio file in a continuous space where distance reflects perceptual similarity, allowing us to compare how classical and learned embeddings behave during nearest-neighbor retrieval.

FAISS Similarity Search

Similarity search is implemented using Facebook AI Similarity Search (FAISS). We use FAISS to allow for fast nearest-neighbor searches on our learned timbre embeddings. At query time, the system returns the closest entries in the index based on Euclidean or cosine distance. Speed is an important factor here, as we need to query against tens of thousands of audio samples. Figure 7 shows a result from the PCA embedding space for an example query.

Figure 7

FAISS search-by-example result showing the top retrieved timbres for a query sound

Note: The model returns clips that share perceptual timbre characteristics with the query, and the 0.9+ similarity scores reflect how well the embedding organizes related sounds.

345