While working my way through the TinyML (Machine Learning models used as part of embedded systems) sequence of courses, I noticed that in courses #2 and #3, the Keyword Spotting (KWS) program did not run. I was a little disappointed to be honest. The courses (offered by Google and Harvard on the edX.org platform) seemed nearly flawless. I was not comfortable completing the course without this code running correctly. I set out to come up with a solution, share it with other classmates, and hopefully apply the technique to some embedded projects I have in mind.
I also was intrigued by the general approach of using neural nets as part of a pattern recognition technique for dealing with time series data. In the past, I have focused exclusively on embedded designs that utilized formulas based on clean, highly deterministic mathematical equations based on experimental data. Instead, these techniques utilized a non-deterministic based mathematical method for dealing with not as well behaved data.
The solution can be found at my github page here.
Note that this requires the first Jupyter notebook be run on Google Colab while the second notebook is run locally on a Fedora install. I have tested it on both Fedora 37 and Fedora 38. I have not tested it on Redhat nor on CentOS, but it should work on any RHEL-based distro.
If there is interest, I will look at rewriting the code as a single Jupyter notebook to be run on a Debian based fork, most likely Linux Mint.
I received a thank you note from Brian Plancher, who is now an Assistant Professor of Computer Science at Columbia University, for my contribution. That is very kind of Brian and I appreciate it!
If you are interested in KWS and want to check out the code, please do!