Estimating Novelty

Synopsis: We can estimate the local novelty in a signal by looking at the accumulated difference of some point in time and its neighbours (argh)

Assumptions: That the measure of similarity is meaningful for the data at hand, that he neighbourhood size bears some resemblances to the time scale of interest

Computing Similarity

  • We're often dealing with multidimensional data
  • Choice of ways of estimating similarity, different trade offs (which are?)
  • Can use different features (e.g. STFT magnitudes or MFCCs)

Summarising Novelty

  • Reduce matrix to a single running estimate
  • Slide a window (the kernel) over the main diagonal of similarity matrix
  • Multiply kernel by what it covers and sum
  • Peak picking (link)