Handwriting Synthesis

The below model generates text that looks like handwriting. Please feel free to try it out

Unfortunately, the university has shut down support for dynamic content. I am looking for a way around this, but I have attached a few samples below for the time being. If you wish to test the model on some new input, please feel free to contact me. I have also removed the option to model my handwriting for security purposes.

The Satrapy

The aim of this project is to produce realistic handwriting that closely matches the user's own handwriting. One of the most widely known papers about this problem is Alex Graves' 2013 paper. It decribes a model that consists of 3 LSTM RNNs, with an attention mechanism and a Mixture Density Network. We can upgrade some of the mechanisms decribes within due to advances since the paper release, but I found much more success with a model that more closely resembles a C-RNN-GAN, as proposed in Olof Mogren's 2016 paper. We keep many of the ideas from the Graves paper and implement them in the GAN framework. There were a host of problems to solve when using a GAN for this (ex. working discrete spaces) which I gloss over here but would be more than happy to detail over an email.