You are here: Home Manual Chapter 10: Similarity Measurements Features normalization for similarity scores

Features normalization for similarity scores

For any new species, or for any novel type of comparison within a species you should normalize your features accordingly:

Go to Options->Features normalization

For Similarity scores you need to normalize feature values per frame:

normalization.png

You can calculate those values using Explore & Score -> Features vectors across interval

You should properly sample your data (e.g., select 20 songs randomly), outline the songs and click Add Rec. You can then copy and paste those records to Excel or Matlab and compute the median values and mad (Median Absolute deviation from the median).

Important: do not include silent intervals in those data, or make sure those are filtered out before you compute the median and mad values.

For DVD maps and for clustering you should repeat the same procedure at the syllable level. Here we recommend you use the feature batch to collect the data in a syllable table, and then from that table you can use Matlab or Excel to compute median and mad values.