Proceedings of the 15th Asia-Pacific Conference on Communications (APCC 2009)-057

$*OREDO&RYDULDQFH0DWUL[%DVHG3ULQFLSDO
&RPSRQHQW$QDO\VLVIRU6SHDNHU,GHQWLILFDWLRQ



&KDJQZRR6HR-XQ<RXQ<RXQJHXQ.LP.HXQKR6LP-DHNZDQ.R

0HLKXD=KDR-RQJNHXQ.LP+HHDH.R(XQ\RXQJ.LPDQG<RXQJKZDQ/LP


7KH*OREDO6FKRRORI0HGLD6RRQJVLO8QLYHUVLW\
6DQJGR'RQJ'RQJMDN*X6HRXO.RUHD
7HO)D[
^ FZVHR MXQH KLQDNL RSHQVKRZP\DXVWLQPHHKZDMRQJQLKHHDHNH\\KOLP`#VVXDFNU
   



Abstract-Principal component analysis (PCA) is an +RZHYHUZKHQFDOFXODWLQJWKHGDWDIXOOFRYDULDQFHPDWUL[
efficient feature extraction method which reduces the WR ILQGW KHHL JHQYDOXHP DWUL[DQGWK HHLJ HQYHFWRUP DWUL[
dimensions of the feature vectors and removes the
correlation among them, with little loss of information, by
RIHDF KV SHDNHU ERWKW KH 3&$I X]]\P L[WXUHP RGHODQ G
projecting the original feature space into a small subspace WKHORFDO 3&$ EDVHG *00UHTXLUH D ODUJH VSHH FK
through a transformation. However, it requires a larger GDWDEDVH+RZHY HU XQOLNHVS HHFKUHFRJ QLWLRQLWLV
amount of training data when calculate the full covariance LPSRVVLEOHIRU DQ\*00 EDVHG VSHDNHUUHFRJQLWLRQ
matrix of each speaker. This paper proposes an efficient PHWKRGWRUHFHLYHHQRXJKGDWDRQDQ\JLYHQVSHDNHU
global covariance matrix-based PCA for speaker
identification. The proposed method uses training data from ,QWKLVSDSHUWRRYHUFRPHWKHVHSUREOHPVZHSURSRVH
all speakers to calculate the global covariance matrix and DJOREDOFRYDULDQFHP DWUL[ED VHG 3&$IRUVSHD NHU
then uses this matrix to find the eigenvalue matrix and the YHULILFDWLRQ7 KHS URSRVHGD OJRULWKPXVHVW UDLQLQJGDW D
eigenvector matrix to perform PCA. During training and IURPDOOVSHDNHUVWRFDOFXODWHWKHJOREDOFRYDULDQFHPDWUL[
testing, our method uses PCA coefficients that are based on DQGWKHQXVHVWKLVPDWUL[WRILQGWKHHLJHQYDOXHPDWUL[DQG
global covariance instead of the PCA coefficients of each
speaker. Compared to the conventional PCA and the WKHFRU UHVSRQGLQJHLJ HQYHFWRUP DWUL[WR  SHUIRUP3&$ 
Gaussian mixture model (GMM) methods, the proposed 7KHUHIRUHGXULQJWUDLQLQJDQGWHVWLQJRXUDOJRULWKPXVHV
method shows better performance while requiring less 3&$F RHIILFLHQWVWKDW DUH  EDVHGRQ JO REDO FRYDULDQFH
storage space and complexity. LQVWHDG RIW KH3&$F RHIILFLHQWVRIHDF K VSHDNHU 7KH
Keywords: speaker identification, principal component analysis, WUDQVIRUPDWLRQSUREOHPFDXVHGE\WKHODFNRIWUDLQLQJGDWD
Gaussian mixture model, global covariance. LV VROYHGE\ XVL QJ WKHQ RUPDOL]HGHLJHQYHFWRUP DWUL[
REWDLQHGWKURXJKJO REDOFRYDULDQFH7 KHH IIHFWLYHQHVVRI
WKH SURSRVHG PHWKRGLVVKRZQWKUR XJKW KHUHVX OWV RIWKH 
I. INTRODUCTION
FRPSDUDWLYHH[SHULPHQWVZLWKERWKWKHFRQYHQWLRQDO3&$
'LPHQVLRQUH GXFWLRQ LVD QHFHVVDU\ VW HSL QW KH DQG*00PHWKRGV
HIIHFWLYHDQ DO\VLVR IP DVVLYHK LJKGLPHQVLRQDO GDWDVHWV 7KHUHP DLQGHUR IW KLV SDSHU LVR UJDQL]HGD VI ROORZV
,WP D\E HW KHP DLQ REMHFWLYHLQWK HDQ DO\VLVIR U 7KHQH [WVHFW LRQ, , SUHVHQWV3& $ EDVHGR QD JOREDO
YLVXDOL]DWLRQRIWKHKLJKGLPHQVLRQDOGDWDRULW PD\EHDQ FRYDULDQFHP DWUL[DQ GVHFWLRQ  ,,, SUHVHQWV WKH*00 IRU
LQWHUPHGLDWHVWHSWKDWHQDEOHVVRPHRWKHUDQDO\VLVVXF KDV VSHDNHUL GHQWLILFDWLRQUHV SHFWLYHO\7KH H[SHULP HQWDO
FOXVWHULQJ 7KHSUL QFLSDOF RPSRQHQWDQ DO\VLV 3& $  UHVXOWVD UH GHVFULEHGL QVHFW LRQ, 9DQ GF RQFOXVLRQVDUH 
PHWKRG EDVHGR QW KHI XOOF RYDULDQFHP DWUL[ IRUV SHDNHU GUDZQLQWKHILQDOVHFWLRQ
UHFRJQLWLRQK DVEHH QL QFUHDVLQJO\ SURSRVHG >@ 3&$
UHGXFHVWKH GLPHQVLRQ RIW KHIHDWXUH YHFWRUVDQG UHPRYH II. THE GLOBAL COVARIANCE BASED PCA
DQ\FRUUHODWLRQDPRQJWKHPZLWKOLWWOHORVVRILQIRUPDWLRQ
E\SU RMHFWLQJWK H RULJLQDO IHDWXUHV SDFHLQWRD VPDOO :HDVV XPH WKDW xs t  R N LV DQ NGLPHQVLRQDO
VXEVSDFHWKURXJKDWUDQVIRUPDWLRQ>@ IHDWXUH YHFWRURIW KHWUDL QLQJV SHDNHU sI RUYHULILFDWLR Q
5HFHQWO\WRUH GXFHWUDLQLQJWLPHDQGVW RUDJHVSDFHD DQGW KDW X S ^X s  s L S ` ZKH UH X s ^xs t  
3&$ IX]]\P L[WXUHP RGHO>@ D QGD*D XVVLDQP L[WXUH
PRGHO *00 EDVHG RQORFDO3&$>@ZHUHSURSRVHGIRU t  L  Ts ` LVDVHWRISWUDLQLQJVSHDNHUVZLWKOHQJWK Ts 
VSHDNHULGHQWLILFDWLRQ7KH3&$IX]]\P L[WXUHPRGHOZDV 7KHUHIH UHQFH YHFWRU PG WKDWL VXVH GI RUIL QGLQJW KH
GHULYHGI URPW KHF RPELQDWLRQ RIW KH3& $ DQGW KHI X]]\ JOREDOFRYDULDQFHPDWUL[RIWKHWRWDOWUDLQLQJGDWDFDQEH
YHUVLRQ RIW KHP L[WXUHP RGHOZL WK GLDJRQDOF RYDULDQFH REWDLQHGE\
PDWUL[7KH*00EDVHGRQORFDO3&$ILUVWS DUWLWLRQVWKH
 S
­ Ts
½
GDWDVSDFHL QWRVHYH UDOGL VMRLQWHGUH JLRQVE\ YHFWRU
TXDQWL]DWLRQ 94 DQGWKHQSHUIRUPV3&$RQHDFKUHJLRQ
PG
S
¦ ®
 ¯ Ts
¦x s t ¾   
s t  ¿ 



‹,(((
 A Global Covariance Matrix based Principal Component Analysis for Speaker Identification

1H[W ZHF RPSXWHGW KH JOREDOFRYD ULDQFH PDWUL[IR U )LJVK RZVWKHWUDQVIRUPDWLRQRI3&$EDVHGRQWKH

WKHIHDWXUHYHFWRUVE\WKHIROORZLQJHTXDWLRQDV JOREDOFRYDULDQFHLQWZRGLPHQVLRQDOVSDFH,Q)LJWKH 
>
6 G ( x  PG x  PG _ x  R N 
T
@ VPDOOHOOLSVHVH[SODL Q WKHFRYDULD QFH PDWULFHV
^6  6   L 6 S ` SUHVHQWLQJW KHGLVWUL EXWLRQ RIWK HI HDWXUH
 S ­ T T½ YHFWRUE\HDFKVSHD NHUIRUWKHFRQYH QWLRQDO3&$D QGWKH
¦ ® ¦ xs t  P G xs t  P G ¾    ELJHOOLSVHLQFOXGLQJDOORIWK HVP DOOHUHOOLSVHV SUHVHQWV
s

S s  ¯ Ts t  ¿
WKH JOREDOF RYDULDQFHP DWUL[ 6 G IURPWKHWUDLQLQ J GDWD
ZKHUH T LVW KHWUDQVSR VH RIWK HP DWUL[7K H JOREDO 7KHUHIRUHWKH WUDQV IRUPDWLRQFD Q REWDLQV XIILFLHQW GDWD
FRYDULDQFHP DWUL[HVWLP DWHGIURP HT  ZDV  XVHGW R UHJDUGOHVVRIWKH XWWHUDQFH OHQJWKG XHW RWKH XVH RI
SHUIRUP3& $W RIL QGW KHHL JHQYDOXH PDWUL[DQ GW KH WUDLQLQJGDWDIURPDOOVSHDNHUVLQFDVHVZKHUHWUDLQLQJGDWD
FRUUHVSRQGLQJHLJHQYHFWRU PDWUL[7KH  JOREDOF RYDULDQFH DUHODFNLQJ
PDWUL[ 6 G RI X S LVJL YHQE\ W KHHL JHQYDOXH
GHFRPSRVLWLRQ>@RI
III. GMM FOR SPEAKER IDENTIFICATION
6 G <G / G <GT    
A. Gaussian mixture model (GMM)
ZKHUH <G LVWK HP DWUL[ RIHLJ HQYHFWRUVDQ G /HW y t  R K EH D KGLPHQVLRQDOIHDWXUH YHFWRU
/G diag OG   OG    L  OG  N LVWK HG LDJRQDOP DWUL[RI
WUDQVIRUPHGE\ WKHJOR EDOFRYD ULDQFHP DWUL[EDVH G3& $
HLJHQYDOXHV6 LQFHW KHL PSRUWDQFH RID Q\W UDQVIRUPHG $*D XVVLDQ PL[WXUHGH QVLW\LVGHILQH G E\DZHLJ KWHG
FRRUGLQDWHLVPHDVXUHGE\WKHPDJQLWXGHRILWVHLJHQYDOXH OLQHDUFRPELQDWLRQRIMFRPSRQHQWGHQVLWLHVDV>@
RQO\WKH K N t K SULQFLSDOHLJ HQYHFWRUVWKHILUVWEHLQ J M

WKHRQHDVVRFLDWHGZLWKWKHODUJHVWHLJHQYDOXHZHUHWDNHQ p y t _ T ¦i 
wi bi y t   
7KHF RUUHVSRQGLQJHL JHQYHFWRUP DWUL[WRW KH RUGHUHG
HLJHQYDOXHVLVJLYHQDV ZKHUHWKHFRPSRQHQWGHQVLWLHV bi y t DUHGHILQHGE\K
: ><  <  L  < @    PXOWLYDULDWH*DXVVLDQIXQFWLRQRIWKHIRUP
G G  G GK
 ­  ½
H[S® y t  P i 6 i y t  P i ¾ 
T
7KHQWKHHL JHQYHFWRUP DWUL[ ZDVXVHG WRWUDQVIRUP  WKH bi y t 
IHDWXUHYHFWRUWRRSWLPDOO\UHSUHVHQWWKHRULJLQDOGDWD S
K 
6i ¯  ¿
'XULQJWUDL QLQJDQ GWHVWLQJ WKU RXJK WKHJO REDO 
FRYDULDQFHPDWUL[EDVHG3&$WKHLQSXWIHDWXUHYHFWRU IRU
ZLWKPHDQYHFWRU P i DQGFRYDULDQFHPDWUL[ 6 i IRUWKHiWK
WKH*00RIHDFKVSHDNHULVWUDQVIRUPHGDV
PL[WXUH7KHPL[WXUHZHLJKWV wi IXUWKHUPRUHVDWLVI\WKH
y s t :TG xs t    
FRQVWUDLQW ¦i  wi  & ROOHFWLYHO\WKHS DUDPHWHUVR I
M

ZKHUH :TG LV D K u N ZHLJKWP DWUL[W KHURZV RI ZKLFK

VSHDNHU¶VGHQVLW\PRGHODUHGHQRWHGDV T ^wi  P i  6 i `i  
M
DUHW KH K SULQFLSDOHLJH QYHFWRUV7 KHLQ IRUPDWLRQ UDWLR
= G JLYHQE\ 7RHVWLP DWH SDUDPHWHUVIURP DOOVHTXH QFHV RI T
K
YHFWRUV Y ^y t  t  L  T ` WK H *00OLN HOLKRRGFDQ 
¦ OG i EHZULWWHQDV
=G i 
N
   T

¦ OG i p Y _ T – p y t _ T   
t 
i 

UHIOHFWVW KHL QIRUPDWLRQUDWH  RIWUD QVIRUPHGGDWD  YLQWKH ,QJ HQHUDOWKHP D[LPXPO LNHOLKRRG 0/ HVWL PDWLRQLV
RULJLQDOGDWD>@ XVHGWRHVW LPDWLRQWKH SDUDPHWHU T ZKLFKPD[LPL]HVWKH
OLNHOLKRRGRIWKH*006LQFHWKH*00OLNHOLKRRGRIWKH
QRQOLQHDU IXQFWLRQL VL PSRVVLEOHW KDWP D[LPL]HVGL UHFWO\
WKH 0/ HVWLPDWLRQV FDQEHS RVVLEOHE\ XVL QJ WKH
x
6 H[SHFWDWLRQPD[LPL]DWLRQ (0 DOJRULWKPLWHUDWLYHO\>@
6S
6 7KHEDVLFLGHDRIWKH(0DOJRULWKPLVEHJLQQLQJZLWK
6
6
6G DQLQLWLDOPRGHO T WRHVWLPDWHDQ HZPRGHO TÖ VXFKWKDW
6 y p y _ TÖ t p y _ T 7KH QH ZP RGHOW KHQ EHFRPHVW KH
y LQLWLDO PRGHO IRUW KHQ H[WLWHUDWLR QDQGW KHSUR FHVVLV
UHSHDWHGXQWLOVRPHFRQYHUJHQFHWKUHVKROGLVUHDFKHG
x
B. Speaker Identification (SI)
 )RUV SHDNHUL GHQWLILFDWLRQ HDFK RI SVSHDNHUV LV 
)LJ7 UDQVIRUPDWLRQRI WKHJOR EDOFRYDU LDQFHEDVHG3&$LQWZR
UHSUHVHQWHGE\*00V T  L  T S UHVSHFWLYHO\7KHREMHFW
GLPHQVLRQDOVSDFH RIVSHDNHULGHQWLILFDWLRQLVW RILQGWKHVSHDNHUPRGHOWKDW


 Proceedings of the 15th Asia-Pacific Conference on Communications (APCC 2009)-057

KDVW KHP D[LPXPO RJO LNHOLKRRGUDW LRI RUD JL YHQ

WUDQVIRUPDWLRQIHDWXUHVHTXHQFHDV
T
sÖ PD[ ¦ ORJ p y t _ T s   
d s d S
t 

IV. EXPERIMENTAL RESULTS

7R HYDOXDWHWKH SH UIRUPDQFHRIWKHSURSRVH G
DOJRULWKP * &3&$*O REDO& RYDULDQFHEDVH G 3&$ 
VSHDNHUL GHQWLILFDWLRQH [SHULPHQWVZHUH SHUIRUPHGE\
XVLQJERWKWKHFRQYHQWLRQDO3&$DQG*00 PHWKRGV,Q
WKHH[SHULPHQWVWKHVSHHFKGDWDEDVHFRQVLVWHGRI.RUHDQ
VHQWHQFHV VSRNHQE \ VS HDNHUV  PDOHV 
IHPDOHV GX ULQJW KUHH UHFRUGLQJVHVVL RQV  UHFRUGLQJV
SHUV SHDNHU RIWKHWZRVH QWHQFHVIL YHUHFRUGLQJV SHU 
VHFWLRQ RYH U WKUHHZHHNV 7 KHJL YHQ .RUHDQ VHQWHQFHV
)LJ7 KH UHODWLRQVKLSEHWZHHQVSH DNHULGHQWLILFDWLRQSHU IRUPDQFHDQG 
XVHGE\ DO O VSHDNHUVZH UH³ yel-lora chamke R SHQ WKHQXPEHURIPL[WXUHV ³mugunghwa kkochi pier-submida” 
VHVDPH ´DQG ³mugunghwa kkochi pier-submida URVHVRI
VKDURQK DYH EORVVRPHG ´ ZLWKDQ DY HUDJH OHQJWKRI  FDWLRQSHUIRUPDQFHZLWKDQGORZHUFRPSOH[LW\
DQGVHF RQGVUHVSHFWLYHO\$PRQJWKHV SHHFKGDWD RYHUW KHF RQYHQWLRQDO3& $  D QG *00  P HWKRGV
XWWHUDQFHVZHUHXVHGIRUWKH3&$FRHIILFLHQWVDQGWUDLQLQJ UHVSHFWLYHO\7K HFR QYHQWLRQDO3&$  P HWKRGOHGWR D
ZKLOHWKHUHPDLQLQJILYHXWWHUDQFHVJHQHUDWHGLQWK HILQDO KLJKF RPSXWDWLRQDOF RPSOH[LW\EXW W KHL PSURYHPHQWR I
VHVVLRQZH UHXVH GIR U WKH WHVWLQJ 7KH VSHHFKGDW DZHUH  HIILFLHQF\GXHWRWKHUHPRYDORIWKHF RUUHODWLRQZDVDERXW
VDPSOHGDW N+ ] ZLWK  ELWVDQGH DFKIUDP H RYHUWKH*00  PHWKRG
OHQJWKZDV  PV ZLWK  PV RYHUODS7KHIHDW XUHYHFW RU )LJ V KRZV WKHSH UIRUPDQFHXVL QJW KHV HQWHQFH RI
FRQVLVWHGRIPHOIUHTXHQF\FHSVWUDOFRHIILFLHQW 0)&&  ³yel-lora chamke RS HQVHVDP H ´ZLWK VK RUW XWWHUDQFH
DQGGHOWDFHSVWUXPLQFOXGLQJGLIIHUHQFHHQHUJ\>@ OHQJWK RI VHF RQGV )URPW KHIL JXUHW KH SURSRVHG
7KHQ XPEHU RI SDUDPHWHUVU HTXLUHGL QW KHS URSRVHG *&3&$   PHWKRGV KRZHGD QL PSURYHPHQWR ID ERXW
*&3&$F RQYHQWLRQDO 3&$DQ G*0 0P HWKRGVL V D QG F RPSDUHG ZLWKW KHFR QYHQWLRQDO
SUHVHQWHGL Q7 DEOH ( YHQW KH SURSRVHG *&3&$QHH GV 3&$  D QGW KH*0 0   PHWKRGVUHV SHFWLYHO\ZKH Q
H[WUD N u K VWRUDJH IRUWK HWU DQVIRUPDWLRQP DWUL[EX WLW  WKH RUGHU RI IHDWXUHYH FWRU ZDVW KHVDP H7 KHS URSRVHG
UHTXLUHV IHZHUS DUDPHWHUVWKDQ  GRWK HFRQY HQWLRQDO3& $ *&3&$   PHWKRGV KRZHGD QL PSURYHPHQWR ID ERXW
RU WKH* 00) RU H[DPSOHWK H SURSRVHGDO JRULWKP S  D QG L QL GHQWLILFDWLRQ SHUIRUPDQFH RYHUW KH
 M   DQG K   U HTXLUHV  SDUDPHWHUV FRQYHQWLRQDO3& $  DQ G*0 0  P HWKRGV
+RZHYHUWKHFRQYHQWLRQDO3&$ S M D QGK UHVSHFWLYHO\%XWWKHFRQYHQWLRQDO3&$  VKRZHGZRUVH
 DQGWKH*00 S M DQGN  UHTXLUH SHUIRUPDQFHW KDQW KH*0 0  P HWKRG E\W KHO DFNR I
DQGSDUDPHWHUVUHVSHFWLYHO\ WUDLQLQJDQG WHVWG DWDRY HU P L[WXUHVGXHWR WK HVK RUW
XWWHUDQFH
7DEOH&RPSDULVRQUHVXOWVRIWKHFRPSXWDWLRQDOFRPSOH[LW\

Methods Parameters
Proposed GCPCA S u M K    N u K
Conventional PCA S M K    N u K
GMM S u M N   

,QILUVWH[SHULPHQWWKHXVHGVHQWHQFHLV³Pugunghwa
kkochi pier-submida URVHV RIVKD URQKD YHEO RVVRPHG ´
ZLWKVX IILFLHQWG DWDOHQ JWK)LJ  VK RZVWK HVSHDN HU
LGHQWLILFDWLRQ SHUIRUPDQFHRI WK H SURSRVHG* &3&$WK H
FRQYHQWLRQDO3&$DQGWKH*00PHWKRGVIRUWKHYDULRXV
QXPEHUVR IP L[WXUHV7KH  GLPHQVLRQVRI  WKHIHDW XUH
YHFWRUVIRUWKHSURSRVHGPHWKRGKDGL   DQGK 
DQGW KHF RQYHQWLRQDO *00 DQG3& $KD G L  )URP
WKH ILJXUHW KHS URSRVHG *&3&$  P HWKRGV KRZHGDQ
LPSURYHPHQWRIDERXWDQGFRPSDUHGZLWKWKH
FRQYHQWLRQDO 3&$  D QGW KH *00  P HWKRGV 
UHVSHFWLYHO\ZKH Q WKHRUGHU RIIHDWXUHYHFWRU ZDV WKH
)LJ7 KH UHODWLRQVKLSEHWZHHQVSH DNHULGHQWLILFDWLRQSHU IRUPDQFHDQG 
VDPH7KHSUR SRVHG *&3&$  P HWKRGV KRZHGD  WKHQXPEHURIPL[WXUHV ³yel-lora chamke´ 
UHGXFWLRQRIDQGDQLPSURYHPHQWRILQLGHQWL


 A Global Covariance Matrix based Principal Component Analysis for Speaker Identification

$OWKRXJKWK H SURSRVHGDO JRULWKPUHTX LUHG H[WUDVWHS V REFERENCES

WRR EWDLQW KHJO REDOFR YDULDQFHEDVH GW UDQVIRUPDWLRQ >@ <$ULNL67DJDVKLUDDQG01LVKLMLPD³6SHDNHUUHFRJQLWLRQDQG
PDWUL[IURPWKHWUDLQLQJGDWDWKHFRPSXWDWLRQDOFRVWVDQG VSHDNHUQRU PDOL]DWLRQE\SURM HFWLRQWRVSHDNHUVXEVSDFH´
VWRUDJHV SDFH IRU ERWKWUDL QLQJDQGWHVWLQJZH UHUH GXFHG International Conference on ASSPSS
GXHW RW KH SUHYLRXVO\FDO FXODWHG JOREDOF RYDULDQFH EDVHG >@ ,7-ROOLIIHPrincipal component analysis6SULQJHU9HUODJ
3&$F RHIILFLHQWV )URPWKHVH UHVXOWVWKH SURSRVHG >@ < /HH- / HHD QG.< /H H³ 3&$)X]] \ 0L[WXUH0RGHO I RU
DOJRULWKPLVHIILFLHQ WIR UUHVR OYLQJWK HW UDQVIRUPDWLRQ 6SHDNHU,GHQWLI LFDWLRQ´ Springer-Verlag, Lecture Note Computer
Sciences 2690SS
SUREOHPFDXVHGE\WKHODFNRIWUDLQLQJGDWD
>@ &6HR .< /HHDQG- / HH³*0 0EDVHGRQORFDO3&$IRU 
6SHDNHU,GHQWLI LFDWLRQ´ Electronics LettersYRO QR 
SS
V. CONCLUSIONS >@ *+*ROXEDQG& )9DQ /RDQ0DWUL[&RPSXWDWLRQVThe Johns
,QWKLVSDSHUZHSURSRVHGDQHZWHVWGDWDJHQHUDWLRQ Hopkins University Press, 3rd ed.
PHWKRG XVLQJDVDP SOHVKL IWRIVW DUWLQJ IUDPHIRUW KH >@ %1)OXU\³&RPPRQ3ULQFLSDO&RPSRQHQWVLQN*U RXSV´JASA
YROQRSS'HF
VHTXHQWLDOSURE DELOLW\UDWLRWHVW 635 7 LQ DX WRPDWLF
>@ '5H\ QROGVDQ G5 5RVH³5REXVWWH[W LQGHSHQGHQWVSHDNHU
VSHDNHU YHULILFDWLRQ $69 7KH 6357L VDQH IIHFWLYH LGHQWLILFDWLRQXVLQJ*DXVVLDQPL[WXUHVSHDNHUPRGHOV´IEEE Trans.
DOJRULWKPIRUUHGXFLQJFRPSOH[LWLHVEXWLWLVQRWSURSHULQ on SAPYROQRSS
WKHVK RUWO\XW WHUHGVH QWHQFHRIV SHDNHU 7KHS URSRVHG >@ $'HPSVWHU1/DLUGDQG''RXELQ³0D[LPXPOLNHOLKRRGIURP
PHWKRGFDQDF KLHYHWKH6357UHJDUGOHVVRIWKHXWWHUDQFH LQFRPSOHWHGDWDYLDWKH (0DOJRULWKP´J. Royal Stat. Sco.YRO
OHQJWKRIWKHWUDLQDQGWHVWGDWDEHFDXVHLWFDQJHQHUDWHWKH SS
QHZWHVW GDWD WKURXJKWK HVDP SOHVK LIWR I VWDUWLQJ IUDPH >@ 6 <RXQJ*( YHUPDQQ'.HU VKDZ*0 RRUH- 2GHOO' 
2OODVRQ'3RYH\99DOWFKHYDQG3 :RRGODQGThe HTK Book
)URPWK HH[SHULP HQWDOUHVX OWVLWLVR EVHUYHGWK DWW KH
(for HTK version 3.2)&D PEULGJH8QLYHU VLW\ (QJLQHHULQJ
SURSRVHGDO JRULWKPVKR ZVW KHUHVXO WV RI JRRGH IILFLHQF\ 'HSDUWPHQW
RIDERX W DQG LQ  (55WK DQW KDW RIWK H 
FRQYHQWLRQDO PHWKRGD QG 6357 EXWO HDGVDOLWWOH 
FRPSOH[LW\ 


ACKNOWLEDGMENT 
7KLVZR UN ZDVVXS SRUWHG E\WK H6RRQ JVLO8Q LYHUVLW\ 

5HVHDUFK )XQG$ OVRZ H ZRXOGOLN HWR W KDQN . </HH

IRU KLV RULJLQDOF RQWULEXWLRQVW RW KH VSHHFKVL JQDO

SURFHVVLQJ :HZLOOUH PHPEHU KLPIR UHYHUDOWK RXJK KH 
ZDVJRQH