/ / / Key word (word spottig), / /
( )
00% DARPA2 0% k 20k 5k % 98898999099992993994995 9969979989992000
Age 90 80 70 60 50 40 30 20 0 0 0 50, 000 00, 000 50, 000 Hours Estimated amout of speech a huma beig hears as a fuctio of age (R.Moore, EuroSpeech, 2003) Word Error Rate(%) 70 60 50 40 30 20 0 0 0 0 00, 000 0, 000 00, 000 Hours Supervised, Usupervised Usupervised (reduced LM traiig) Extrapolated word error rates for icreasig quatities of traiig data(r.moore, EuroSpeech, 2003),000,000
0 5 0 4 00 30 60 20 40 word jucture rule bottom up top dow
X K L X N, L, X L X K N {} {} S (t) S' ( t) r ( t) M rm ( t) X M X N d M d K Ci S(t) :, S (t) : r(t) : M X : N di : Ci
PARCOR 5 0 20 FFT 20 5 20 FFT 8 4 MFCC 8 2 LPC 8 4 8 2 8 2 3 8 2 3 6
LBG = = = = = = (0) (0) 0 ) ( 0,, 0, },, {, }, 0, ; { ) ( D m y y A N j x N j ε L L ( ) { } = = = < = 0 0 ) ( ) ( ) ( ) (,. ), ( ), (, }, 0, ; { } { (2) N i j i j m i j m i j m t j m i j i N m j S x y x d D S x y x d y x d t N i S N A x L ( ). 4., /, (3) ) ( N m m m m A D D D ε < { } { } ( ). 2,,.,, (4) ) ( ) ( ) ( 0 ) ( + = = = + + + + m m S C y y y A i m i m N m N m L ( ) ( ) = = = 0 0, arg mi,,, ˆ i x i x d x x x C x L
2 ( ) = M = A (2) A y i (3) A A ( x, x, L ) 0, = C 0, 0, M x { y, y, L, y }., { y0, y0 +, y, y +, L, ym, ym + } A = { y, y, L, y }. 0,2M = 0 + y 0,2M 0,2M i 0 M 2M { y ; i = 0,,, 2M } y i,. M, LPG = L i M = 2M 2. = N.
(a) (b) VCV SPLIT (c) ( )
/ / ) ( )
{} {} jucture rule {} {} {} bottom up
DP
3
The little boy ra quickly The little boy ra quickly 2 3 4 5 6 7 the 8 a 9 little 0 big boy 2 girl 3 ra 4 walked 5 quickly 6 slowly the a little big boy girl ra walked quickly slowly
JUMP
4 78 6 2345678 (a) 4 7 8 4 2345678 (b) (DP) 2345 2345 3245 3245 (c) 2345 6 2345 6 5555 6 (d) 5555 6
I+JCI I+JC(I-) k I J-k)!/I-k)!(J-k)!k!
D( A, B) = mi F w( k) d ( i( k), w( k) j( k)) b J C x = ( I, J ) w( k) = i( k) i( k ) + j( k) j( k ) j = i + r warpig fuctio b j C = ( i, j) w ( k ) = i ( k ) i ( k ) C 4 C 2 j = i r C b 5 2 C 3 C D( i, j) b = (,) D ( i, j) = mi D( i, j ) + d( i, j) a a 2 a i ai D( i, j 2)
DPDTW) 2 2 2 g( i-2, j-) + 2 d(i-, j) + d(i, j) g(i, j) = mi g( i-, j-) + 2 d(i, j) g( i-, j-2) + 2 d(i, j-) + d(i, j) j i (i,j) g( i-2, j-) + d(i-, j) + d(i, j) g(i, j) = mi g( i-, j-) + d(i, j) g( i-, j-2) + d(i, j) g( i-2, j-) + d(i, j) g(i, j) = mi g( i-, j-) + d(i, j) g( i-, j-2) + d(i, j-) + d(i, j)
b J D ( i, J ) R b 2 b a 2 a i ai B ( i, J ) i = u( j) test patter 0.5 0.5 (a) (b) (c) (d) (e) test patter Asymmetric DP path ad weight for word spottig (base axis : referece patter)
. 2. 3. 4. 5. D (, i) = D (0, j) = ; execute 3.4.5.6. for = execute 4.5. for = D ( i,) = D ( i,) B ( i,) = i for =,2, L N j =,2, L J j = 2,3, L J ˆ i = arg mi D ( i', j ) i 2 i' i D ( i, j ) = D ( iˆ, j ) + d ( i, j ) B i,2, L I,2, L N ( i, j ) = B ( iˆ, j ) Iitializatio Iitializatio for word boudary
Example 36 4 2 4 02 2 3 4 d : I N J D : I N J cumulative distace local distace 03 03 3 4 2 3 0 0 2 2 2 3 46 22 00 0 2 2 2 3 2 2 2 2 2 2 0 0 4 4 test patter
A = a a2 a I L B = b b 2 LbJ pa qa r A s A A B
2A=abcdef B=bdcgfh p=q=r=s= A = a, a2 L ai, B = b, b2 Lb J DPDPDP DP /3 3 A B
DP )