API examples using 8-mile for audio

The codebase relies primarily on 8-mile (mead-layers) for its modeling and optimization code. Whats left is pretty much just training and inference code

The code depends on:

• editdistance (for error evaluation)
• numpy
• six
• soundfile
• mead-baseline
• pytorch

There are a few optional dependencies

• scipy (for on-the-fly resampling of wav files)
• ctcdecode (for prefix beam decoding with optional LM)