LANGUAGE MODELS FOR AUTOMATIC SPEECH RECOGNITION OF CZECH LECTURES

This paper describes improvements in Automatic Speech Recognition (ASR) of Czech lectures obtained by enhancing language models. Our baseline is a statistical trigram language model with Good-Turing smoothing, trained on half billion words from newspapers, books etc. The overall improvement from adding more training data is over 10% in accuracy absolute, while using advanced language modeling techniques mainly neural networks yields another 3%. Perplexity improvements and OOV reduction are discussed too.