THESIS
2005
x, 70 leaves : ill. ; 30 cm
Abstract
During the past decades, automatic speech recognition technology has been applied in different commercial sectors around the world, including China. However, Chinese speech recognition systems are often built un-der the framework for English, despite the fact that Chinese is a language which is in many ways different from English in terms of linguistic charac-teristics. The goal of this thesis is to exploit the linguistic characteristics of Mandarin Chinese, including tone and character-based writing, for Chinese speech recognition....[
Read more ]
During the past decades, automatic speech recognition technology has been applied in different commercial sectors around the world, including China. However, Chinese speech recognition systems are often built un-der the framework for English, despite the fact that Chinese is a language which is in many ways different from English in terms of linguistic charac-teristics. The goal of this thesis is to exploit the linguistic characteristics of Mandarin Chinese, including tone and character-based writing, for Chinese speech recognition.
Tones play an important linguistic role in Chinese. Although different methods have been proposed to integrate tones for Mandarin recognition, there is yet any convincing measure to quantify its importance. In this thesis, an innovative way of quantifying the importance of tone for Chinese speech recognition is introduced. Furthermore, a new approach of incorporating a tone classifier into a state-of-the-art recognition system will be given as well.
Characters are the basic units in the Chinese writing system while the definition of a word is controversial. For speech recognition, new words often are hard to predict or capture. For Chinese, new words are always decom-posed into a sequence of characters and the character set is fairly static. This property is exploited when we augment our language modeling data by extracting text from the World Wide Web. Instead of extending the vocab-ulary size, new words are implicitly captured via the contextual information between characters.
Experiments on conversational Mandarin speech recognition showed that the proposed tone classifier integration and the use of web data for Chinese language model are useful in improving Mandarin recognition accuracy.
Post a Comment