- For a dataframe of audio data:
- each column is a sample
- each row is a point in time/amplitude of a sample
- So when using for machine learning,
- each column name is a label
- each column series should be converted to a row of features (eg: mean, max, std, tempo mean, beat mean, stft mean etc)
- you can also use percentile to get value at specific percentile and use them as features
- NEVER SHUFFLE (time series like index will be lost)
- Missing data should be interpolated for time series
- Rolling window can be used along with stat functions as well for generating new features