Summary: 

Goal of this paper is to automate the process of labeling scratch
bouts from a time series of audio recording data in a mouse. You use
two steps method to analyze the data.

The first step is to segment the data into 'potential' scratches and
remove any observations that is clearly not a scratch. For this
purpose you rely on the fact that series of scratches for a scratch
bout show certain rythmic behavior. You have outlines the steps for
segmentation clearly.

The second step in the analysis is then to build a classifier to label
the data selected from the first step. For classification you will use
random forests classifier. You will also build various common
features, extracted from the data (which could be transformations of
the observed data), and use them for classification. I think feature
design needs some more work in this writing.

Specific comments are in the attached pdf.