A novel statistical language modelling method for continuous Chinese speech recognition

Statistical language models can play an important role in continuous speech recognition, but their performance is often unstable because of the training data sparsity. This paper proposes a statistical language modeling method, where the contribution of the language model is limited by the acoustic matching result and the N-gram probability distribution is modified referring to the length of the silence duration between adjacent syllables. Besides, the paper proposes a powerful single-state hidden Markov model (HMM) to model various kinds of silence segments