公共关系学 2-居延安 著 复旦大学出版社【企业营销策划经典】

1、MODULATION SPECTRUM EQUALIZATION FOR ROBUST SPEECH RECOGNITIONSource: Automatic Speech Recognition & Understanding, 2007. ASRU. IEEE Workshop onAuthor : Liang-che Sun , Chang-wen Hsu, and Lin-shan LeeProfessor : 陳嘉平Reporter : 楊治鏞IntroductionnThe performance of speech recognition systems is very

2、often degraded due to the mismatch between the acoustic conditions of the training and testing environments.nIn this paper, we propose a new approach for modulation spectrum equalization in which the modulation spectra of noisy speech utterances are equalized to those of clean speech.IntroductionnTh

3、e first is to equalize the cumulative density functions (CDFs) of the modulation spectra of clean and noisy speech, such that the differences between them are reduced.nThe second is to equalize the magnitude ratio of lower to higher components in the modulation spectrum.Modulation spectrum (1/2)nGiv

4、en a sequence of feature vectors for an utterance, each including D feature parameters,nwhere n is the time index, and is the parameter index. ( ),1,2,.,x n nN( ) ( ,1), ( ,2),., ( ,) , 1,.,Tx nx nx nx n DnN1,.,dDModulation spectrum (2/2)nThe modulation spectrum of the -th time trajectory can be obt

5、ained by applying discrete Fourier transform:( )dY kd10( )( ) exp(2/)NddnY kynjnk N1,2,.,dD0,1,2,.,1;kNSpectral Histogram EqualizationnWe first calculate the cumulative distribution function (CDF) of the magnitudes of the modulation spectra, , for all utterances in the clean training data of AURORA

6、2 to be used as the reference CDF, .nFor any test utterance, the CDF for its modulation spectrum magnitude, , can be similarly obtained as . ( )dY krefCDF ,( )d testYktestCDF Spectral Histogram EqualizationnHence the equalized magnitude of modulation spectrum is1,( )( )d testreftestd testYkCDFCDFYk,

7、( )d testYkMagnitude Ratio EqualizationnWe first define a magnitude ratio (MR) for lower to higher frequency components for each parameter index as follows:nwhere is the cut-off frequency used here, is the order of the discrete Fourier transform.d0 120( )( )ckdkdNdkY kMRY kckNMagnitude RationWe can

8、observe from this figure that the mean value of is degraded when SNR is degraded, and thus is highly correlated with SNR.nIt is therefore reasonable to equalize the value of for a noisy utterance to a reference value obtained from clean training data.dMRdMRdMRdMRMagnitude Ratio EqualizationnWe first

9、 calculate the average of for all utterances in the clean training data of AURORA 2 as the reference value .nWe then calculate the value of for each test utterance as .dMR,d refMRdMR,d testMRMagnitude Ratio EqualizationnWe equalize the magnitude of the modulation spectrum for the test utterance as w

10、here is the weighted-power for the scaling factor.,( )d testYk01p,c,c,(1),()( ) ,kk( )1( ) ,kk()d refpd testd testd testd testd refpd testMRYkMRYkYkMRMREXPERIMENTAL SETUPnAURORA 2nThe speech features were extracted by the AURORA WI007 front-end.nFigure 4Integration of MRE with Other FeatureNormalization TechniquesnWe only consider MRE her