摘要:  Recurrent neural network language models (RNNLMs) are becoming increasingly popular for a range of applications including speech recognition due to their inherently strong generalization performance. However, two issues limit its application for speech recognition systems.

   The first one is the decoding problem. N-Best rescoring was used for RNNLMs instead due to its intrinsic characteristics. In this talk, two novel lattice rescoring methods for RNNLMs are introduced. The first uses an n-gram style clustering of history contexts. The second approach directly exploits the distance measure between hidden history vectors.

   The second problem is the computational cost during training. It is hard to parallel derived from the long-term history characters. Previous researches mainly focus on class based output layer and distributed asynchronized training on multiple CPUs. This talk describes an alternative approach that allows RNNLMs to be efficiently trained on GPUs with bunch mode.

 

报告人简介:  Xie Chen is a third year PhD student in the Machine Intelligence Laboratory from Cambridge University, supervised by Prof. Mark Gales. Before joining in Cambridge University Engineering Department in 2012, Xie obtained Bachelor degree at Xiamen University in 2009 and finished M.Phil, studied at Tsinghua University (in Electronic Engineering) in 2012, supervised by Prof. Jia Liu. His Research interests mainly lie in machine learning and speech recognition, especially for decoder for LVCSR systems, acoustic and language model adaptation, deep neural networks for acoustic modelling and recurrent neural networks for language modelling.

 

联系人:  张卫强，62787115