Neural network language models for conversational speech recognition

Recently there is growing interest in using neural networks for language modeling. In contrast to the well known backoff ngram language models (LM), the neural network approach tries to limit problems from the data sparseness by performing the estimation in a continuous space, allowing by these means smooth interpolations. Therefore this type of LM is interesting for tasks for which only a very limited amount of in-domain training data is available, such as the modeling of conversational speech. In this paper we analyze the generalization behavior of the neural network LM for in-domain training corpora varying from 7M to over 21M words. In all cases, significant word error reductions were observed compared to a carefully tuned 4-gram backoff language model in a state of the art conversational speech recognizer for the NIST rich transcription evaluations. We also apply ensemble learning methods and discuss their connections with LM interpolation.