Recurrent Neural Network Language Modeling Applied to the Brno AMI/AMIDA 2009 Meeting Recognizer Setup

This paper is on Recurrent Neural Network Language Modeling Applied to the Brno AMI/AMIDA 2009 Meeting Recognizer Setup.

In this paper we use recurrent neural network (RNN) based language models to improve our 2009 English meeting recognizer originated from the AMI/AMIDA project, which to date was the most advanced speech recognition setup of the Speech@FIT. On the baseline setup using the original language models we decrease word error rate (WER) from 20.3% to 19.1%. When language models in the system are replaced by models trained on a tiny subset of the original language model data, WER drops from 22.2% to 20.4%. Adding data sampled from two RNN models for language model training improves the overall system, yielding the performance of the original baseline (20.2%).