Preliminary exploration into automating conversion from audio files into Christmas / Halloween light shows. Datasets, weights, and other large files are excluded from this repository due to their size.

This is a very slow and cumbersome process to do manually, and I am hoping to find a way to generalize it. My Dad loves his Christmas lights, and I want to save him time and energy.

My early efforts have been unsucessful so far due to the complexity of the problem. I initially believed that audio feature extraction would be sufficient detail to learn light show conversion, but there are patterns and dependencies between units that cannot be captured in a singular time unit.

• Use only Unit 1 as output (simplify)
• Create converter for compressed files (gain access to many more files) - https://github.com/Cryptkeeper/lightorama-protocol
  • Include Halloween songs as well to maximimize dataset (underlying principles should be the same)
• Conduct more rigorous correlation examination and exclude features that appear to have little correlation with sequence
  • In my brief initial examination, there seemed to be roughly half of the extracted features that had essentially zero correlation with any of the units
• Include some amount of previous sequence information in training input
  • I.e. n previous centiseconds output + n previous centiseconds audio + this centisecond audio
  • Domain knowledge would imply that the previous sequence choices are very important for current sequence choice