We previously published an article that applied the
Hanning-3 segmentation method in conjunction with HMM:
http://class-specific.com/csf/papers/Vehicle.pdf to the
Office Sounds data set (http://class-specific.com/os/).
The script software/hanning3_officesounds.m implements the classifier mostly as described.
The script software/officesounds.m with mrhmm=0
runs the experiment using just 17 training events per class,
for all 6 holdout sets. The results in that paper can be
reproduced. We obtained 81 errors in 12240 opportunities,
or 0.66% error.
In addition, at the end of the script software/officesounds.m,
a code segment is provided that allows time-series data to be
synthesized from the trained parameters.
Examples of real and synthetic time-series are shown in Figure 12.4.
The time-series (as .wav files) can be accessed here:
software/keys_real.wav, software/keys_syn.wav,
software/stapl_real.wav, software/stapl_syn.wav,
software/sciss_real.wav, software/sciss_syn.wav.
Figure 12.4:
Example spectrograms of real data (top)
and synthesized data (bottom) for Office Sound data classes
from left to right: "keys", "stapl", "sciss".
|