Summary: | 碩士 === 國立暨南國際大學 === 電機工程學系 === 93 === In recent decades, more and more researchers are dedicated to developing the techniques of automatic speech recognition. In order to apply the speech recognizer in a real environment, the mismatch between training and application conditions must be handled carefully. For example, various kinds of additive noise in the application environment often degrades the performance of the speech recognizer seriously. In this thesis, we focus on the problem of additive noise and use several approaches to alleviate its influence on speech recognition. These used approaches are categorized into two classes, speech enhancement and the robust speech representation.
The objective of speech enhancement approaches is to modify the testing speech features obtained in the noisy environment and make them better match the clean conditions for the pre-trained models. In chapter 3, four approaches of this kind are studied. They are Energy Contour Enhancement (ECE), Spectral Exponent Adjustment (SEA), Spectral Subtraction (SS) and Wiener Filtering (WF).
One the other hand, the robust speech representation can reduce the sensitivity of the speech features to a mismatched acoustic condition, and thus decrease the mismatch between the clean training and the noisy testing environments. We study four approaches in deriving robust speech representations in Chapter 4. they are Autocorrelation Mel-Frequency Cepstral Coefficients(AMFCC), Phase Autocorrelation(PAC), Cepstral Mean Subtraction(CMS) and RelAtive SpecTrAl(RASTA).
In Chapter 6 and 7, the recognition experiments and discussions are presented. The experimental results that these approaches can effectively improve the performance of the speech recognition under the noisy environment. Furthermore, in order to test whether these approaches are additive, we combine ECE or SEA with all the others to process the speech signals. Experimental results show that such a combination of two approaches can provide better recognition accuracy than the individual one in most cases. As a result, we can conclude that these robustness approaches reduce the influence of additive noise in different directions and are thus additive.
|