Using neural networks on raw audio waveforms is hard, because they contain so many datapoints (tens of thousands per second). Therefore, we usually transform the waveform to an image - the spectrogram. The spectrogram is then used as input for the neural network to process sound. However, it's somewhat tricky to again recover an actual sound waveform from a spectrogram because some information is lost in the transformation process. In neural_vocoder.ipynb, I'll train a convolutional neural network (CNN) to infer cat meow waveforms from spectrograms. The aim is to use a simpler architecture than cutting edge models, whilst still giving decent results.
