Comparison of Various Neural Network Language Models in Speech Recognition

In recent years, research on language modeling for speech recognition has increasingly focused on the application of neural networks. However, the performance of neural network language models strongly depends on their architectural structure. Three competing concepts have been developed: Firstly, feed forward neural networks representing an n-gram approach; Secondly, recurrent neural networks that may learn context dependencies spanning more than a fixed number of predecessor words; Thirdly, the long short-term memory (LSTM) neural networks can fully exploits the correlation on a telephone conversation corpus. In this paper, we compare count models to feed forward, recurrent, and LSTM neural network in conversational telephone speech recognition tasks. Furthermore, we put forward a language model estimation method introduced the information of history sentences. We evaluate the models in terms of perplexity and word error rate, experimentally validating the strong correlation of the two quantities, which we find to hold regardless of the underlying type of the language model. The experimental results show that the performance of LSTM neural network language model is optimal in n-best lists rescore. Compared to the first pass decoding, the relative decline in average word error rate is 4.3% when using ten candidate results to re-score in conversational telephone speech recognition tasks.