标题: | 利用加速演算法之大词汇连续语音辨识系统 A Fast Large Vocabulary Continuous Speech Recognition System |
作者: | 杜明桓 Tu, Ming-Huang 陈信宏 Chen, Sin-Horng 电信工程研究所 |
关键字: | 大词汇连续语音辨识;Large Vocabulary Continuous Speech Recognition System;LVCSR |
公开日期: | 2008 |
摘要: | 本论文主要探讨如何建构大词汇连续语音辨识系统的加速演算法及其应用。第一部份先对语音辨识系统各部份的加速演算法做一系列的研究与实做分析,主要由声学模型、语言模型和搜寻演算法三方面去做加速,而使得辨识率的下降在极小的范围内。在Treebank语料实验中,可减少50%以上辨识时间,其辨识率几乎无衰退;在非特定语者TCC300语料的实验中,辨识时间可降低22~44%左右,而辨识率只下降在1%范围内。 其次,第二部份对于辨识系统建立不同的应用,设计出有文法规则的辨识系统,并且使该系统更具弹性。可藉由简易的方式去调整系统的模型与参数,方便往后使用者去实做出不同需求的语音辨识系统。 This thesis can be divided into two parts. In the first part, large vocabulary continuous speech recognition (LVCSR) by speedup algorithms is constructed. The thesis describes some effective algorithms that reduce the computation of the acoustic model (AM) , language model (LM) and search space. In the outside of Treebank, the system recognition speed can be accelerated by more than 50%, and maintain the same recognition accuracy; besides, in the TCC300, the recognition speed also can be accelerated by more than 22%, and the character accuracy just decreases by less than 1%. Therefore the system is capable of the speaker independent recognition. In the second part of the thesis, a flexible LVCSR for the different applications is built. The user can not only tune up the system’s parameter and on line, but also be easy to design the grammar-ruled word net, and compile the language model which the recognition system can read in. |
URI: | http://140.113.39.130/cdrfb3/record/nctu/#GT079613563 http://hdl.handle.net/11536/41999 |
显示于类别: | Thesis |
文件中的档案:
If it is a zip file, please download the file and unzip it, then open index.html in a browser to view the full text content.