Nothing too exciting here yet. Some small experiments I've been working on to build up to algorithmic composition using RNNs and largely-unsupervised learning directly from waveform training data.

Likely to be a fool's errand.

Currently consists of:

• A iPython notebook prototyping out some FFT transforms and MEL visualizations.
• A C# project that captures the Windows audio loopback device and allows you to easily annotate music you're listening to with beat onsets to build up a respectable training set.

• Currently using a Windows host
• Anaconda2 installed in C:\Anaconda2, 64-bit, Py2.7 (source)
• VS2013 installed in default folder for complilation, VC folder added to path for pycuda operation.
• Git installed and added to path
• Followed instructions for CUDA7.5, mingw, libpython, theano, pycuda (source)
• Jupyter notebooks started from Anaconda Command Prompt
• pip install pyglet for multimedia playback
• git clone https://github.com/jameslyons/python_speech_features; cd python_speech_features; python setup.py install for MFCC lib