N-gram language models in JLASER neural network speech recognizer

In our recent research we have discovered that neural networks can be more efficient in speech recognition than the state of the art approach based on Gaussian mixtures. This statement is valid only for small corpora, however, many applications do not require a huge recognition vocabulary. In this article we describe our speech recognizer — called JLASER — based on neural networks. We also show the effect of n-gram language models applied to the JLASER recognizer.