Summary 

The researchers collected data of mouse scratches in the form of audio
recordings of a single mouse in its cage.

The goal was to find scratch bouts in the audio recordings. Scratch
bouts are made up of rapid swipes that occur with a period of about 50
ms. If they mice were not having a scratch bout, they were having what
the authors called scratch-free periods.

For each mouse, the authors split each recording into intervals over
time. They did this by searching for a characteristic rhythm of the
swipes in a bout.

Then within each time interval, the authors examined the frequencies
and found that the sound of the swipes has a broad frequency
band. They also found that the higher frequencies were not
contaminated and the lower ones were. So, they wanted to find the
points that had higher frequencies ¡V these are the peaks ¡V and they
wanted to make sure the peaks were about 50 ms apart. If they were,
then this interval contained a scratch bout.

When the authors received the audio recordings, they first created a
spectrogram, which allowed them to visualize the sound waves. They
then ran a Fourier analysis to determine which frequencies were
contained in the waves. Then they were able to find which time points
had frequencies above 10 kHz. For those high-frequency points, then
they were able to measure the power. They then made a new time series
out of the power measurements. They ran a triangular kernel smoother
on the time series (why?).

In the smoothed time series, they found the points that had maximal
power within 25 ms and power greater than the minimal time bin within
25 ms by some chosen parameter h. (I don't understand this part.) They
found the sequences of time points that had between-adjacent-point
distances under 120 ms, and marked as candidates the ones that had at
least three time points.

The values selected for the cutoff points in the analysis above were
determined from exploratory data analysis (do they agree with other
authors?).

It wasn¡¦t clear to me whether the sections about classification,
transformations, and common features are expanding on the breakdown
from the segmentation section, or whether they were a different
process altogether.

Comments 

This project is very interesting. Although it seems useful for
detecting mouse scratch bouts, it could also be used as a methodology
to detect patterns in other sound waves and even images. The author
could sell this methodology angle a little better. So far, it seems
difficult to find other experiments that would match the ¡§mouse
scratching¡¨ one, so it might be dismissed by those outside that field,
but it could actually be very useful outside the field.

It wasn't clear to me what features were determined by the data (from
exploratory data analysis) and which were determined by the
scientist. This should be made clearer.

It is common to use the past-tense in the methods section. This might
make it a little more cohesive. Right now I wasn¡¦t sure which parts
are ¡§usually done¡¨ and which were actually done.