Loading...

Introducing a Hybrid Language Model for Improving Performance of Continuous Speech Recognition Systems

Bahrani, Mohammad | 2010

712 Viewed
  1. Type of Document: Ph.D. Dissertation
  2. Language: Farsi
  3. Document No: 41084 (19)
  4. University: Sharif University of Technology
  5. Department: Computer Engineering
  6. Advisor(s): Sameti, Hossein
  7. Abstract:
  8. The utilizing language model is one of the most effective methods for improving speech recognition performance. For speech recognition applications, several types of language models have been proposed for speech recognition applications that try to model some parts of language information, such as n-gram models, syntactic models, and semantic models. Although n-gram, syntactic and semantic models are able to model different structures that exist in natural language, they each only capture specific linguistic phenomena. None of them can simultaneously take into account all of language phenomena in a unified probabilistic framework. Recently, a number of semantic models called "latent topic models" attract attentions. In this research, we intend to propose a combination of latent topic models and n-gram model. Our goal is the relaxation of "bag of word" assumption fundamentally present in latent topic models. First, latent topic models are surveyed and then different methods of combining language models are studied. Among latent topic models, PLSA and LDA models are chosen for combination with n-gram model. The graphical model based method is used as combination method. Two new hybrid language models named NPLSA (n-gram and PLSA combination) and NLDA (n-gram and LDA combination) are proposed. In NPLSA model, the independence assumption between latent topics and context words that present in previous methods is relaxed. For both models, inference and parameter estimation procedures and also mathematical formulation of hybrid models are extracted by standard methods. It is showed that the proposed models are generalization of PLSA and LDA models. Both models are trained using part of BLLIP WSJ corpus and evaluated by perplexity and word error rate criteria. The results show that the proposed models outperform the PLSA and LDA models and previous hybrid models. The special case of NPLSA model for n=2 (BiPLSA model) improve perplexity about 20% than the previous (Nie et al's) method. The perplexity reduction of NLDA model compare to LDA model is about 24%
  9. Keywords:
  10. Language Modeling ; Language Models Combination ; Latent Topic Models ; Probabilistic Latent Semantic Analysis (PLSA) ; Latent Dirrichlet Allocation (LDA)

 Digital Object List

  • محتواي پايان نامه
  •   view

 Bookmark

No TOC