Maser s Thesis A sudy on realizaion of speech and speaker recogniion sysem based on feedback of recogniion value (Kim, Hyun Goo) Deparmen of Mechanical Engineering, Division of Mechanical Engineering Korea Advanced Insiue of Science and Technology 2005
A sudy on realizaion of speech and speaker recogniion sysem based on feedback of recogniion value
A Sudy on realizaion of speech and speaker recogniion sysem based on feedback of recogniion value Advisor: Professor Min-Yang Yang By Hyun Goo Kim Deparmen of Mechanical Engineering, Division of Mechanical Engineering Korea Advanced Insiue of Science and Technology A hesis submied o he faculy of he Korea Advanced Insiue of Science and Technology in parial fulfillmen of he requiremens for he degree of Maser of Engineering in he Deparmen of Mechanical Engineering, Division of Mechanical Engineering. Daejeon, Korea 2005. 1. 8. Approved by Professor Min-Yang Yang Advisor
. 2004 12 23 ) ( ) ( ) (
Speech recogniion is he process of recognizing wha is spoken and speaker recogniion is he process of idenifying who is speaking by he some informaion included in speech signal. Speech recogniion echnology is usually used many kinds of man-machine inerfaces and speaker recogniion echnology can be used o verify speaker s ideniy and conrol access o services such as voice dialing, banking by elephone, elephone shopping, daabase access services, securiy conrol for confidenial informaion areas, and remoe access o compuers. This hesis proposes realizaion of speech and speaker recogniion sysem based on feedback for sabilizing various recogniion values depend on spoken speech samples and speakers. Mos of speaker recogniion sysem is realized as ex-dependen ype o ge high recogniion efficiency, bu speaker recogniion sysem realized in his hesis have o be exindependen ype because spoken speech sample is recognized simulaneously. Because exindependen ype can change recogniion value depend on spoken samples and speakers, recogniion using sysem realized in his hesis is more difficul han he ex-dependen sysem. So, in his hesis, compensaion on he basis of feedback recogniion value is used o solve he problem which is resuled from realized sysem.
H(z) 1 = 1 αz
α α
g() x() h()
x() = g() * h() = g( τ)h( τ) dτ 0 X( ω) = G( ω)h( ω) X( ω) G( ω) H( ω) x() g() h() log X( ω) = log G( ω) + log H( ω) log X( ω) log G( ω) log H( ω) c( τ) = F 1 log X( ω) = F 1 log G( ω) + F 1 log H( ω) c( τ)
c( τ) = F 1 log X( ω) = F 1 log H( ω) c i [n] [ c [n] c [n] c [n]... c [n ] T c (n) = 1 2 3 p ]
c1 c 12 η
c () i = c i [n] η K k= K kc [n + k] i
H ( z) = 1 0.95z 1
[ q j q = i, q = k, q = m,...] = P[ q = j q i] P = 1 2 3 1 = [ = j q i] a P q 1 = ij = a ij
λ = { A, B, Π} Π λ λ Pr( O λ1) Pr(O λ ) λ λ Pr( O λ) 1
= = = = = + + + + + = ξ = γ ξ = γ β α β α = λ = = = ξ 1 T 1 1 T 1 N 1 j N 1 i N 1 j 1 1 j ij 1 1 j ij 1 j) (i, (i) j) (i, (i) j) ( ) (o b (i)a j) ( ) (o b (i)a ) O, j s i, P (s j) (i, : Define ) arg max P(O ˆ : For λ λ = λ i γ γ = = k,o j (j) (j) (k) bˆ converges ) ( unil Repea ˆ λ λ λ X P = ) ( ˆ) ( λ λ X P X P ξ ξ = j ij j) (i, j) (i, â : Reesimae
Define : δ (i) = Iniializaion : 1 Ψ (i) = 0 Recursion : δ (j) = max{ δ 1 (i)a Ψ (j) = arg max{ δ max S,S,...S δ (i) = π b (o 1 i i 2 i i 1 1 P[S S ) 1 1 2 ij 1...S }b 1 j (o (i)a } ij S ) = i,o λ] Terminaion : P* = max{ δ ( i)} s * T i = arg max{ δ ( i)} i T T Backracking : * s ( s * = Ψ + 1 + 1 )
) (o b (j)a (i) Recursion ) i,s o o o Pr(o (i) Define i j ij 1 3... 2 1 + β = β λ = = β (i) ) Pr(O i = β 1 λ (i) ) Pr(O S F i T α = λ ) i,s,...,o,o Pr(o (i) Define 2 1 λ = = α ) (o b (i)a (j) Re cursion j i ij 1 α = α ) (o b (i) Iniializaion 1 i 1 π i = α
Define λ = (A, B, π ) Pr( O S, λ) = b s 1 (o 1 )b s 2 (o 2 )...b s T (o T ) Pr( S Pr( O,S Pr( O λ) = π S a a... a 1 S1S 2 S2S3 ST 1S T λ) = Pr(O S, λ) Pr(S λ) λ) = Σ all_s Pr(O S, λ) Pr(S λ) = Σ T Π a all_s = 1 S 1 S b S (o ) T TN N 2 T
4 11