For visualization purposes, plot the beats on top of a waveform of the song

Currently, memory usage is an issue when using multiple datasets. Some things that may help:

• Saving features for individual files instead of the whole dataset
• Using pescador generators to sample from these files

Define interfaces for other beat tracking and tempo estimation datasets

Because the number of frames can be so large, the inputs to the networks may be huge if we present the entire song to the model at once. LSTMs are known not to capture very long term dependencies very effectively, so we may be able to get away with running backprop on parts of songs instead of the whole song at once. We'd have to look at how we can train the model effectively so that the hidden states make sense in between time steps. Perhaps we can run the network on the input sections in both directions to get the proper hidden states and outputs for the song, and then run backprop on each section independently.

Investigate the effect of using stronger prior smoothing (e.g. larger k for add-k smoothing, or larger Gaussian filtering) for reducing the potential for the prior to select an guess that was not salient according to the model