Recurrent neural network language models (RNNLMs) have recently shown to outperform the venerable n-gram language models (LMs). However, in automatic speech recognition (ASR), RNNLMs were not yet used to directly decode a speech signal. Instead, RNNLMs are rather applied to rescore N-best lists generated from word lattices. To use RNNLMs in earlier stages of the speech recognition, our work proposes to transform RNNLMs into weighted finite state transducers approximating their underlying probability distribution. While the main idea consists in discretizing continuous representations of word histories, we present a first implementation of the approach using clustering techniques and entropy-based pruning. Achieved experimental results on LM perplexity and on ASR word error rates are encouraging since the performance of the discretized RNNLMs is comparable to the one of n-gram LMs.

Index Terms: Language model, recurrent neural network, weighted finite state transducer, speech decoding