Search

A new unit, named word fragment of language model was proposed to take full advantage of the Chinese linguistic information among adjacent syllables, and an algorithm for word fragment selection was studied.The experimental results show, with the language model based on word fragment, syllable accuracy for recognizer is improved and the speech retrieval system gives better performance than the one with only syllable based model.

The traditional method of speakers segmentation based on Bayesian information criterion（BIC） was improved.New speakers segmentation method adopted a pre-segmentation process to decrease calculation spending.Theory analy-sis and experiments show that the method of speakers segmentation based on pre-segmentation makes the amount of cal-culation controlled,on the premise of realizing the basic segmentation performance.

Syllable lattice based Chinese speech retrieval methods can avoid the problem of out of vocabulary (OOV) words and compensate the retrieval performance loss resulted by recognition error. For absence of effective indexing method in lattice based retrieval approaches, a posterior probability based indexing method is proposed in this paper, which introduces syllables and K step neighbor syllable pairs as index items and takes the posterior probability as weighted value for an improved vector space model. It is proven by a series of retrieval experiments that our method is more suitable for lattice based spoken document retrieval tasks and the improvement accomplishes its anticipated purposes. [ABSTRACT FROM AUTHOR]