Paper Reading: COVER DETECTION USING DOMINANT MELODY EMBEDDINGS

Problems to solve: automatic cover detection.

Method: a neural network architecture. Each track is represented as a single embeded vector.

Question:

• what is embedding extraction?

Notes

• Cover detection is not a classification problem. It requires the comparison between tracks.
• Gerenal Ways:
  • Compare:
    • Dominant melody
    • Tonal Progression: a sequence of chromas
    • Chords
    • Fusion of both
  • Compute similarity score:
    • cross-correlation
    • a variant of the DTW algorithm
    • combination of both
  • Problems:
    • Good on small datasets, expensive with the expansion of the dataset.
• Faster Methods:
  • comparison of all possible subsequences pairs between 2D-DFT sequences derived from CQTs overlapping windows.
  • Encode track as single scalar or vector: the comparison work can be done offline. Just need to compare the Euclidean distance between embeddings.

Words & Phrases

• Stricto sensu: 严格意义上