A model takes possibly multiple inputs and produces possibly multiple outputs. In Nabu a model is trained for a specific task by minimizing some loss function. Every model in Nabu is an encoder-decoder model. An encoder-decoder model consists of an encoder that encodes the input into some other hidden representation and the decoder decodes this representation into the desired output.

You can find more information about encoders here and decoders here.

An example of such an encoder-decoder model is the Listen, Attend and Spell model that encodes the input to a high-level representation with a pyramidal DBLSTM (Listener) and decodes this high level representation into text characters with an attention-based rnn-decoder (Speller).

However, not every model can as easily be viewed as an encoder-decoder system. For example a DBLSTM trained with CTC cannot be devided into to parts in a straightforward way. In Nabu this system is simply implemented with a DBLSTM encoder and a linear decoder that maps the encoder output to the desired output dimension.

Similarily, an rnn language model is implemented with a dummy encoder that does nothing and an rnn decoder that predicts the next character based on the history.

You can build new models by combining new encoders and decoders or by creating new encoders and decoders.