Exploiting Topic- and Position-Dependent Language Models for Mandarin Continuous Speech Recognition

[[abstract]]This study investigates language modeling for Mandarin continuous speech recognition. First, a word topical mixture model (WTMM) was proposed to explore the co-occurrence relationship between words, as well as the long-span latent topical information, for language model adaptation. During Speech recognition, the search history is modeled as a composite WTMM model for predicting a newly decoded word. Second, a position-dependent language model was presented to make use of the word positional information within documents and sentences for better estimation of word occurrences. The word positional information was exploited in conjunction with that information provided by the conventional N-gram and probabilistic latent semantic analysis (PLSA) models, respectively. Finally, we also attempted to develop a probabilistic sentence-ranking framework for extractive spoken document summarization, for which the sentence prior probabilities were estimated by the whole sentence maximum entropy (WSME) language model that tightly integrated the extra information clues extracted from the spoken sentences for better selection of salient sentences of a spoken document. The experiments were conducted on Mandarin broadcast news compiled in Taiwan. The speech recognition results revealed that the word topical mixture model and positional dependent language model, respectively, could boost the performance of the baseline large vocabulary continuous speech recognition (LVCSR) system, while the spoken document summarization results also demonstrated that the integration of extra sentence-level information clues through the whole sentence maximum entropy language model could considerably raise the summarization accuracy.