Hybrid Neural-Network/HMM Approach for Out-Of-Vocabulary Words Rejection in Mandarin Place Name Recognition

In this paper we address the problem of rejecting Out-Of-Vocabulary words in speaker-independent Mandarin place name recognition. We integrate neural network and Hidden Markov Models in an attempt to utilize the strength of both. HMM based acoustic models including keyword models, filler models, and an anti-keyword model were trained to meet our needs. Statistical features are fed to a neural network for further verification. Feed-forward backpropagation network, Elman backpropagation network, and trainable cascadeforward backpropagation network were compared. A baseline models based on conventional N-Best normalization methods was built. Experiment results showed that feed-forward backpropagation network achieved the best performance, which reduced average error rate by 54.4%.