Dissertation
Dissertation > Industrial Technology > Radio electronics, telecommunications technology > Communicate > Electro-acoustic technology and speech signal processing > Speech Signal Processing > Speech Recognition and equipment

The Research of Feature Extraction Algorithm for Speech Recognition and the Realization

Author HuiBo
Tutor FengHongWei
School Northwestern University
Course Computer Software and Theory
Keywords Speech recognition Endpoint detection Mel Frequency Cepstral Coefficients Dynamic Time Warping
CLC TN912.34
Type Master's thesis
Year 2008
Downloads 803
Quotes 14
Download Dissertation

The voice signal having a strong time-varying characteristics , and the characteristics of the speech signal in a shorter time interval can be regarded as essentially unchanged , which is an important starting point of the speech signal processing . The level of the speech recognition rate , and also depends on the accuracy and robustness of the voice signal feature extraction . Therefore , the voice signal feature extraction plays a decisive role in the speech signal processing applications . The thesis studied the basics of speech recognition , including speech recognition principle ; basic knowledge of the speech signal processing ; speech recognition and training methods . On this basis, this paper completed work : 1 , focuses on the widely used Mel frequency cepstral coefficients (MFCC) parameters , the 24 -dimensional MFCC parameters , for example , changes in component analysis of the higher order parameter is missing the recognition rate identify a higher order MFCC parameters are sensitive to noise , and in the recognition rate changed little in the case of 24 -dimensional MFCC parameters optimized combination . 2, using the VC model based on dynamic time warping (DTW) to achieve a connection string of digits speech recognition system and experimental analysis . The basic configuration of the composition of the system module and the voice recognition system model . In the realization of selected Mel frequency coefficients (MFCC) . 3 , the experiment found that the Chinese digital easy to confuse the issues , and made ??improvements in the template training methods and reference template is proposed to use multiple pairs of feature vector sequences robust training and acoustic vowel segmentation to construct a reference template method. Finally, this paper studies the Chinese continuous speech recognition acoustic modeling methods identify Chinese easily confused words . This article through experiments and research on the part of the actual speech recognition system for further work to develop practical speech recognition system to do the basic work .

Related Dissertations
More Dissertations