This is a simple neural truecaser written with AllenNLP, and based loosely on (Susanto et al, 2016). They have an implementation here, but being written in Lua, it's a little hard to use.
We provide pre-trained models that can be used for truecasing English and German right out of the box. The English model is trained on the standard Wikipedia data split from (Coster and Kauchak, 2011), and achieves an F1 score of 93.01 on test. This is comparable to the best F1 of (Susanto et al, 2016) of 93.19.