| Summary: | 碩士 === 國立成功大學 === 電機工程學系 === 86 === In this thesis, a Speech Input Query System for Chinese Web Page
is described. The user can query data on web page using speech
input. This query system includes a network part, a language
model part, a speech recognition part and a query part.In
network part, the system can get web page dynamically and build
keyword model of the web page. The language model part deals
with the ambiguity in segmenting sentence by Viterbi algorithm.
It caculates unigram and bigram information from vocabularies of
web pages, and builds common vocabularies of web pages from
query sentence. In the speech recognition part, we use
VenusDictate as the recognition kernel.It first builds
character-lattice and word-lattice using vocabularies and common
vocabularies of web pages, then finds the optimal word sequence
using full search and BIGRAM language model. In addition, we
propose a grammar rule and modifies part of BIGRAM language
model in order to speed up recognition time and increase
recognition rate. In query part, the system offers many syntax
for query, and the result can be displayed on web brower or
output by a TTS system.We use web page on NCKU EE and NCKU IIE
for test, and design 100 query sentences and 100 keywords for
expriments. The recognition rate and recognition time depends on
the size of web page. Experimental results show that the top one
recognition rate is 93.1%, top ten recognition rate is 100%, and
the average recogniton time is 0.23 second.
|
|---|
