From 41d4fa5ff9338bae38240e4ce8b909eb4dfffc20 Mon Sep 17 00:00:00 2001
From: yemaozi88 <428968@gmail.com>
Date: Tue, 5 Mar 2019 00:11:38 +0100
Subject: [PATCH] sp is added to the model.

---
 .vs/acoustic_model/v15/.suo      | Bin 102400 -> 100864 bytes
 acoustic_model/fame_functions.py |  13 ++--
 acoustic_model/fame_hmm.py       | 106 +++++++++++++------------------
 3 files changed, 52 insertions(+), 67 deletions(-)

diff --git a/.vs/acoustic_model/v15/.suo b/.vs/acoustic_model/v15/.suo
index 712c255714ef6ffda4e8bb7262540e56cc859336..96899cb98ddd6e7708d34b29a0a50a9133e53c15 100644
GIT binary patch
delta 6346
zcmd^D4OEm>8vfpo8RTa`K)?ZE91s){5d;)aU<L^dDZ#%`2|)!F2~ojBQAdnbj1cBp
z*VIr)V_i>WG1l~BHmU8pR;%kVt+OrS)*h{O&06)SrR;NO1|(zmZ0&aToZWlQGtd41
zdEfWG@BPNG%h|BSxhaNSUFWLyk|cLYk`%z*&dyGdM1bcjxdUX^oWSz0+E1S$J*$$}
z(sS&(yq=0!wbLq(R`h8P7riCPg^{A)S}k&zptB#?*OABM(jY+TDBnT*VC2<601yNW
z0D^&0fY?PTFc63b!~)SkKHvu|0=$7?z+*rkz#qW0Ydc3EqXN=^p}-j60U#6*N1Fv6
z4fF>hfH)u)a08-%WMBxO15$uMAO?s8FuXGXoDTE_Mgk84nSf~9CAOA<!gyc;@DPv+
zga8i%qSGX>50C@o0+WF$z*Jy5Ao}Dvc({W_T@vz-0tLWqK+~K%EN>z*S%B!U*umVv
zqE4hLM|nk$`jw9SDsT<3wny93$Q$J{*QVq>V};tw?pM3tPW0$fyM2w$P7j;<h1x|;
zQIGjV`H>#w`_(SmyV5VvZy?a6_G09P+C}>Yuu!|mH-LrOMP8Vf(E2zb8&Pq;+C^QR
zqx1c04@3PtK&V~x6X|}n&q13|dwFwizu~UPv;acQ_U`uIo@y2bE|O1k?trzz(nX=r
z!6jg4NB(>IMBNNPSoKU`0U!)^K6oy$5SRuCqZ9LqB#cjRg@c!XmjacJ@^VQUZtL(k
zDykios~s!`inIo0gQL#q=Bs!LH=n895<|r5&!E78t-!N@sN3$~yDT<zP;*5X>u`%O
zcerWE<SGA~6L`r+btUa!RqC~r!mdZHr&?AWu*z*2_IFk&UozkI2$jv+z<o~a_>@GG
zruP8>o6+@S+kku0T93Mo&>_=?Cc27)Y(+(4bM8Q@ZO-i<{411DX-hBk$iImEtLUYE
z4H*mYI`9VYCh!)(E{6swk0Nuox#(}k=t>%Eu8Mw=jDevnRC|^_Ww*57ItN2&r-<Di
z`k+!O7lZUIQLHlZq>_YU?WJLp*f{T4RV})T{flH8tT44IjH*h-aMZFDstk6>N2Q1p
zg%5ldD7J`9qDPG4X^|=OsMORUvl3VZh(oRhuW{rJWWGB>Rv0TX6v9lzs>+dQTg`6y
z4P_>6sEe>9A#B$8#|Y<XX!-(su6S4>oQJ4ml}l`!RBk}q2J@|iKF+2LnVU+;$Ly2p
zDKoAAK85f&qO<v$zlux~WG?GyB-d1t58pI{!q|dsACr;k!_t^xyTOzxw`E-;H?h}F
z43x0P0NF97sf;|$x@@bG<+bR{I-`|LTV-R5fl8S^JcGB)Cl{WbNgA*D(K{^L@6@UO
z{NXvh>Yc|c*t<1;tf0}K6}AMXu7XPuQr-^E1V60BHBmsw;Q2Vm@<DI_;J`5K8|`0%
ze}l3bk!QEW*6gfp%L|ol@(~Wx0}#hkZbP3(S#HZjmfJW8D<S_q4rt52Dbqf7tDv)w
zh#yFtbTryLs_=xX@%3@8B<`i~#dguo%yGeuwZ%bO#kH9>S+mMnHg<2qdgAsXalMZG
zh8gPnf8cvX|3rBD!M&~v9uV%K+k{8LFBUstdhl0RA#LirDebrsW6>wwN0KICvO;hy
z_$6>V>JPzx3&nm0K8*4vKL|sXr1!ufEM?kJ(;<ajnJb<ojX9_6W#UUyDA}A``D-#x
ze{@d5hqd4AzWRr&t%r&`9vbqQ9&>bWC?69EBkt!SN$bR7;1l3}uJ$PX5%?37FS<$6
zCG?wyFm7{f;6I`KI^;f(c=JTkn;xOmbN%6w(ER}7iSSHU!M{U!Q=lZ(DJAJN_(jx@
z3dZ@c^poHjV0Re7Cz3gJ>=80emYb(~nZ0TfX;kU^Z|*H?3Vo$*&)YdZL&smlOx-)r
z#m+)tFXKF6o$%czUr8E_@{j=Yn$@+kap^G^5kzwVp*Q1kh54q{r4O7_t4tIWiAZ%o
zD@i3-y&k4^1G}<$8T_#J%CNPPjEO(O=!g4CQWbbAxD@OT-L1l?r=hn!C~pUE1-}h_
zy^2{{pr<J)-v-YHAIHW<pni(K{Q-0oT!7UtLN?Kmi92L+1npBb*bz8RD@lt<4@ZGa
z4kkDe0)3z%3ENa+;S%sJ)Q8|kB!J7oao`%<jJMG5&R|>YZ41*;&JE-*VNCoGNqQN5
z4oBd);6X6`cJzG&{3gobQIgc?1kVe;fO2dMelTvrNIaIn-J)r_5ylRNXOqlA<bGz~
z<Cm4DEebZQE{ACwVwhotmE|<N%+}O<1+T*lPXOzIjevMJB>}eQ5ewcl03ju5(=gpO
z)NKcZuXzrf2-E_*0kNzJya(6|JP-U55cT`O`vJChQxqL*&P`LkhZ5ZVUrc*tW3AOE
z#+Fvs#6-o^WEPbzD=uas4H`E4jtei;k{3Hx9pN=9Dk0w9xDtcIr!n1ng9<XSJtd(S
ziA~xZ7??n^v}|c~MRggTsO8|Hcr7HyB(I8ZJ9VayM<@D;Bsj`^;oLykOIjiQ79~j7
zhHY@H6=uVy2D%rD*em&i*rKy43S!hA#AbXsotZuv$kv@t+haISrdF9vz8GW<y7D=h
z(&Zl!Y_$J8*x;sIl}%3&)7J+uYrT%ms`K+n9N8tK1=w-;v}MK2P-o!D^GR)1Z;T*j
zsP50-@u2}iXzmG732}Cz6*wf?kG)qt#4Rq$A<NaM9?RQDk~f=B7omzvjP06!ju;?j
zQt=93a$)855t{h;*r-JNe8TmL)tBZhU0kxTtT;S>neizDi@i09m2P%3Tc0uz%c<75
z#m01vU=<B&X4q^%z-yau%}F7w$b9qF$e$7!j0HFCfq^2;x88bR(VGKtYP)77)joyY
zLR|5Q0>h6t-IYx%<yJS6O$M16Zp5n=;{{ucXL5<O7$JSIn(NZ2hM!SUnef#*Zt$Sx
z?sM_P&Bq<DmhN3AH}Rx`>%dc*$Ar;NcJk&sJjRK(!m4?RH*K@Yl`k4e5f)D`8cwED
z3gdU*v1jji+WK>yl00~glCn%~GQX^$=lEu#CEO52k*xgN2+QFRS|}^wfTTc+HJTnG
zZ!cVMdy23IBOkyuF?3QnZd`Kyw6?Qns22n0yVaD$3U4{{^M16!Up%DlPZ>w~?|rF>
zH>sg)Y&V(jP*axggP8Rn+>QhA%e~!&R%cR}t|*L`tb452dKEtR?n^`XPA3}8uMVg8
zEkpa!IcK4c)*yJQe^np8_w*s2sl6znl~;OG!}m2%LL7Q%;7v(1`uke2Knu%=&Hi?@
z5c*>>g>eY{1scI$=}Uuc7eJ1T?k-I9KGZG72&*R);cI~+JjgD><Tx7VvA9Q%)qGPN
zE$3<8lwgU{(k%2#q6%(EqJey1IxNS!QsEAt?nvW>#};*7U3K${%kqonF0Nv#F+oGi
z@qtny72$UQKI|)ClVf4J6)@s*d@g&~zQx5%#XQ%YV%e2T&Q6(93`+~2*?0VF`;V=X
zjzwi(T9aCGK+E*0ewI7&<Rl9t6N$b1*Q7pqC*4GzZ-}Q5R``XQS)VcRuVd)Ae_;2a
z5W<m1G?#j}9$`ZBct;-^`lBV`+I8w~IrQzSaU74ij|u+oh+(Z%kNBL8qjw@cS7Bdc
z5Q8qDd=&BcI>a#+(K8#I4%UM&f=kh_D%SpidJ6mkEaqymB+Wtlb?^+Zb&UPb2@x<w
z>Rqu;QsP|ZG$uUshemMfa!|pRtm4ngp7;26S>C7H@rL*1-T@RM7tQ5{c#2>l->JRY
z--t+mDf7$``aIyok=QS?c~d;8d83AUYf@4L&9Kl|I#2xca?<|aot_njt>D@;a^@W=
zv`>D<79cDL5JAKbE~b_2cA+nHtLFOYw1|i4DU)XM3n`SqPZi?bQWx%vi<JxQmffrE
z{_G0lVdH5(ugIi>R4dw-@Qy6<N8GUGby*b6Cr_kCn$DNZBRx-?NP$#xZz7iTr520l
z@UNyLbYx7Ymw3l4n&-M4|K*5i)l+lam`9axeYTlP^GGYUzlCqvNK3oFEF;BgH9UVd
z`N+oZP`H4fC?J2HUXPv~voSEegr@THBJ$)NrR2sw2vhw)LL!7(xlbiRXjvge(+>N(
zr%K6-*A$Vqhma!e`O(}kha#v|oOC`{%%KcgdGDlr$6N#@+hz+Z5qTOP4dgj<p_L<P
z6v@9Sqg0-%BhAmCl33oBN)v>LTr4^B=rmD5k8R-tS|cvTN*+HSAtSVewnAB5`uAn^
zqg5841vH1cbP5SXivdBra5`o4h*BEGD;H9>T|x<bm!2}%?Z!~cwnenRdsM|n$gQ!G
zG?t9@luTkb(QvL^>v%C`n1<_?LmmgqprylA_Fc8umG=5_3bbjKHKwRwj#)6rCbncu
zF29;Z&RskEsn&+DX4{7~54Qgwu&+0NR%5Xl`~QX}EjKq%773eM%5|IYNWl$la7vbJ
b;UVCM;X#t}pm>LG!L7tCN>M%YBw7Cs2B(ej

delta 5943
zcmd5=3sjWXwVpi;Pk{oC78P+oG|0=D0S2T-2Sr2$10))fn8;HR2_gc;i15cm&4U=6
zljf+2iW9GS*4X*mj!6_p^Kjj^*CuJ2+N>lO?{c-ts%c1SR-;Knzx~gToh0qL*1Otu
z4}NR!bN+MoIcLAl9+gKzl!1_bYm(8Rw-|#V3aknEYGh=DM}vTvA5}~;h7WwlWnS^d
zTT~-DjJ=dCJ~Q4!En;KPrl`G`vu_&DHW<bUGL0R_5^e^@0uZD%^UjK~z47{h0-69!
z0`h=LfM;z4jnH|PxB@5ymI6J%cL2Y*k&(~<ekS;NdjA&CG{6cZ>to;d)83Kqf++#A
z0UIy{V1d``^fsMN)TvXaGeB!~J`Qv~5CSX!{(|+zposv>S^_!;C<P`1_X0c~%h#N)
zcX(hK_}lgVjXF)&X)x$YfDLg6=&e9Ca06fm76EmD=$jB<*bIgRa%&CXxxXI$_JFZY
zozDcFrSso|ZA^>N5nobkxZ(HJr`gi8HA`QwS^6w2-P=2AWtPir43_Zi`H-WHZO-<e
z3YdWj0GllZ2nW~&c#Q%Y=mNZ+m9qh07f1rw1+oC$1$qq~=<viGPz%65z#0n!*axD4
ze1PW!`T!rW4dV{rn=TNEvFiYKfdU`{7!R-uu%?{=YkP)at-&xC%t9azU?1R(*aw)7
z`al`_OMpNh=t937hy{3V1?WvW&;8{(U!~KgofQkKu%F>heV_$&9l&Pe1>D$t-_!eS
zJ^$#)$Xz=B1ZX#~8R*r=egNu<Fq?R;h>yq@C1V=JP(-Ra6k#D(R=bH+CAzYTVweep
z_(fJqFfSVspJmPU&&!`_^)t;g+Y@>6U1Tt90+`<n+N1Ldsb7svHma7Gp}<{IBWGYj
znm$Q9mzCfb_lvB?Yui9roKKlJ1jLSKQ`L*kNOHxsvnR)CIV-L=I31f6h>c@6g>1sK
z6tjAH+?k*l_BxJ|(F?0*C!`u($)D46;!^S#6Ed;rbA0A@H?Re`2e=P`BP@ye<0Lw}
zfX4t<5w`(+%sqes7Rv2E^pClqPXbQ?PXo^Z{uv`9KLr2NofT8fUNA?2mjXVD3yBF)
zuLcafCSo7^gQw3)QnAu4F_V`D9XvZNesAn2uVigI((o&BvU{8ud}y9~y2+I?Kb+W!
zID}X^qCMr8Vt%Ymv`@PsoYr8J1BAZ>Z^3OJ7MEf-kzG@qO?(!24{1t^+=fYs9M@<r
zOFP-s4$Fh&wwhdN7pYTxmexxhq9gr5w<BcEqHzYpCq{#zCEj4z4VjA&Qd@)Z$XKu$
za>Nnjgwq)R3F7_(%p29ghY0DDMs-ftDe}y*z-h)0kvqfao^2FLWU`uE{62YJxfnEN
zGS-gpLC7Gfn6hR6DvDN}rM_Tyyipp9$tj%8K3B7OMtOAS)Ru;i&YyVaFCT7+uy7u_
zvfC%fP7OhDTV$^7;`%#Zc6HvKboy7}f2#S|bmsoqr;g1W_wk<t7rcpWD`4BKy|@b}
zgJ~hiLKwU+)L=NpXZl9d7HCJjUj4GuM0|@UlB<luU05mFo65sDVELuc>~gW`hG60>
zEee-si1xLx4$F(uzX$hYht2Ax+eVD^p15?!NN6^`ps@<_R7dsC!rd*}?d|gD{gkMd
zw;dsOtI?HHWMVI8slgAi40!eo;a*)CvJAcZ4dQanQq%W&hUOi)`o3NINaS|R+Kvu4
zm-0#&hYzPQo~~Y0B1yItQh{2~{u;Tr7)8Zk`WU-CBd5E#rg=j{uc+8HQyxjDsOY+N
zEu9T*?JHXw*R?ck5Xvr}v7sH)GlH@+?Bb!_MPuq4R@HPgx35~$+#udKo?*6TSp8CZ
zu-&v}8yecm+uZjWJ>4Y~v$NuK(IiXlslCHp2VaQWm;skJ$GwAhU@?sH0348w6P=*;
zY{r>}x5lV3+lI-t&nIFBo6XsI8Fq(Xv@$AMIQJ{jtWtiorUEr}!nr#swiypcOIGW~
z8F=>#tOc8L+%J+SJU(9TPp3rraR`Np*j;mjYgvv^UQ%FWi)xuTM4n?#5^wxwq8Q#6
zCJqngMB4N7MlZDKidw|Vk7C5am*OVh4FOlu3;3<T;{dC^PfXr3i5?cS_M{_cn`Uq#
zVc*yTx)0b7xb-m}KdAH1?5y~%cxq4Tgr8vGg@6e>=|%KkM#iC6kW4<_*=iEjUE_l7
z4o8MnEZ*hNj3&x<KjAu|i1oYUBd%DZ3x`iseJ2#?;iE)(Je?xNvcd5}IjN9QBpJu5
z<{cMEq?9L!{>l{9S?MG9BPQW{!69F-pxN@pN{W!y)d<-KZt5hCSG!DHLvb>(l2Sxt
z(OIX}nrRJao4f5fjtpDA-(u`mR_o?&n=N-#oqV0YwU$t!EKr2~;e+bI{uLyXN@<#y
zw){4n#%Fo|h17Y6i{Cm$=O_H^l<mXPKLlvygZw|v71fm%q1>!+j<aexPPiT}Qj1*2
zNDR9VsfQlFmE372_kfm`><0$bLH9gz&qGkVs@3X)1LP^WNG48E>gT7^t}WIhv{=W<
z5M^K8e%T-Il&xO+ub-Q~6@Q`P!ge#OJow51MZYLfO9uKvqxhNO#&m&t?#U@e&!M;I
z>%#)WF@`RK6KBRh4SU4N>8)PHNY4|w1l_br&MBsNQF?fpID0sikWHSS>wca@{qqST
z!IPr;JWldF6=aNhTNl8BszqUUaoRf6FU=^T*I|n+L#tsU(1u=<wo(0j6kjV0%TWI{
zsLC@6d5$7%oF1oFi6U@{$0@9U?`<XWJ-i(po#7}H#T)x_<bg7pA>Xf}7&*O^EI8eA
zIj@-FaJ(gl=iqprv+z;pZxm7KN$@5_(D@+OyFQ`(RuNsE+2YH0@AbUxqi=~7%6p1y
zC?~Q7MOcHO1@*xS+;@7sbS|JqXyfFg9ddRBMF`JPvv_MLV@wL(MLZ&?Mqd`TFZ$&$
zAtIi=kSd=3-~{cJ2OQLd1P~<_U%pN@Cea}=>(f5(%Bi%M=rRA-pC2D1e*I}K@%k$F
zS>q3g@~?=oMH}X2-gpy|57JL|eqLRzEx$g~Vav_V%(3M<GH2Oy^D<}E)Mc-lWw++l
z);4VRK0A%-$o&8U^s-gk@*5jTTRNQ(PyZ<@<!Z(CkQccwHpu<1k#$pug4Bw?LSR%k
z-B3nzpniPk99qS&1`6DuRS$kGX}kX;#YiaT9T9z@Pt;tPBnL8)UsoPCd-LMxKSRZy
z3$Jp@&HFpox#ds-vEJqg=ZC5Pyvih0_NjNkL0=J4a-O&PI=a(Hx$2oqWfWV@-mJGq
z@ZsJ}yl5nb_J>Gk4&67N%c#-D#TK`suRDVrAb93tY0ITW#<WEml_d_EA6pwBoonZ*
z-spT9BxBLnoZ~wAQW!lW4tz1`8qUW*Uq;(1FizlV&leSVz8<;DL7iN)XwBwhi}KI8
zNC3qYH!=C#)X=J5@{c)#IDKxW#&p-zt?StE=v+z_C;uED^UX9YYjVkY$VNtOM$!l0
zo*`l{TZQjY<;mzgGV)I1!2T&>Zdu~BZbek(MTYrrBDY+>mNv++?KEBtTf=0`0va#P
z^^_p{>!?*W7g3z?QO(W0NC62%x5{%%Xn}myN@kp-Lhjr~Q>3+?N@dkTRAW2j`Vz|J
z98n}LSDW!86@@Tzpg6gEitGuc2j$UOWd3{hJvJs#k9QYLzUIrx|5<|FUtK;OwdoDm
z{V<>DA+hvX3r-v)?<}Do%Jw?Eyq+?Oms#bo%f5o?;&<m__@}DDZ!YGLzU9JfO47`x
zTh%c7dXAcnL|bX299~Qzw1t(qUOE>;hxaa_6|!|Lb&E4yk+NqAg~&(B$ih#44y!v-
zI_>29Urqk+G89y$8A`M6)#&cWWmF^|*#<|^jK8#uish*V)FuZ?DM6fB5h905ap$2W
z)GXV}$wi%h&w#p%d0pF;x+g8bX~zFE%B;W2{X_fk{}y;@Wkm(@kISF5hbyT^vybev
z!5FVs;;`qNaIB#X^x~+O2b;+Z&D;!Mc-(K>p&K#pLLI8Ogqz7ihvX;4v|83zQ4!X_
z&GJOviDbChL=KD&zcO~*K2g4Y3;g3Qf9(9RiVFO?Lg^RgEx&`_B$-k}yZnK7q>LuY
zW7X8i4&(5@@r~6exZzR>Qv1s9zSx=<<BhAOT9O_PKLqVQcIyGcQU4Z$zGUy726~3%
z`PFnxx+|zj{;nJ+zbbRw@h{L;og8Rn)o(-T^4e-#3uo4xAxr*v6}Z$UN~C(f9|rd&
z$|Gz1mgx)7PInWugW&3y)m2pT|6KE4Xug2{?-%yAZe>T&oD?a>B+KtZo`K)A`F~an
iZijnl%4}Ya5Ztj5Ddv9iq4amkwzWv+Q2CqJwEVBUko>j)

diff --git a/acoustic_model/fame_functions.py b/acoustic_model/fame_functions.py
index 9ca7e0d..9f4e127 100644
--- a/acoustic_model/fame_functions.py
+++ b/acoustic_model/fame_functions.py
@@ -378,17 +378,22 @@ def ipa2htk(ipa):
 	return ''.join(htk_splitted)
 
 
-def performance_on_stimmen(stimmen_dir, hmmdefs):
-	#hmmdefs = r'c:\OneDrive\Research\rug\experiments\acoustic_model\fame\htk\model_\hmm1\iter20\hmmdefs'
-	#stimmen_dir = r'c:\OneDrive\Research\rug\experiments\acoustic_model\fame\htk\stimmen'
+def performance_on_stimmen(config_dir, stimmen_dir, hmmdefs):
 	lattice_file = os.path.join(stimmen_dir, 'word_lattice.ltc')
 	hvite_scp	 = os.path.join(stimmen_dir, 'hvite.scp')
 	#fh.make_filelist(os.path.join(stimmen_dir, 'mfc'), hvite_scp, file_type='mfc')
 	hresult_scp  = os.path.join(stimmen_dir, 'hresult.scp')
 	#fh.make_filelist(os.path.join(stimmen_dir, 'mfc'), hresult_scp, file_type='rec')
 	lexicon_file = os.path.join(stimmen_dir, 'lexicon_recognition.dic')
-	chtk = pyhtk.HTK(config_dir, fame_asr.phoneset_htk, lexicon_file)
 	
+	# get feature_size from hmmdefs.
+	with open(hmmdefs) as f:
+		line = f.readline()
+		line = f.readline().strip()
+	feature_size = int(line.split(' ')[2])
+
+	chtk = pyhtk.HTK(config_dir, fame_asr.phoneset_htk, lexicon_file, feature_size)
+
 	result = chtk.recognition(
 		lattice_file,
 		hmmdefs,
diff --git a/acoustic_model/fame_hmm.py b/acoustic_model/fame_hmm.py
index 4e2f430..5555919 100644
--- a/acoustic_model/fame_hmm.py
+++ b/acoustic_model/fame_hmm.py
@@ -26,7 +26,7 @@ make_mlf		  = 0
 extract_features  = 0
 flat_start		  = 0
 train_model_without_sp = 0
-add_sp = 0
+add_sp = 1
 train_model_with_sp    = 0
 train_model_with_sp_align_mlf = 0
 train_triphone = 0
@@ -35,6 +35,9 @@ train_triphone = 0
 
 # pre-defined values.
 dataset_list = ['devel', 'test', 'train']
+feature_size = 39
+improvement_threshold = 0.5
+
 hmmdefs_name = 'hmmdefs'
 proto_name   = 'proto'
 
@@ -47,7 +50,9 @@ sil_hed		 = os.path.join(config_dir, 'sil.hed')
 prototype    = os.path.join(config_dir, proto_name)
 
 model_dir    = os.path.join(default.htk_dir, 'model')
-
+model0_dir = os.path.join(model_dir, 'hmm0')
+model1_dir = os.path.join(model_dir, 'hmm1')
+model1sp_dir = os.path.join(model_dir, 'hmm1sp')
 
 # directories / files to be made.
 lexicon_dir = os.path.join(default.htk_dir, 'lexicon') 
@@ -55,9 +60,6 @@ lexicon_htk_asr = os.path.join(lexicon_dir, 'lex.htk_asr')
 lexicon_htk_oov = os.path.join(lexicon_dir, 'lex.htk_oov')
 lexicon_htk     = os.path.join(lexicon_dir, 'lex.htk')
 
-
-#model1_dir	  = os.path.join(model_dir, 'hmm1')
-
 feature_dir = os.path.join(default.htk_dir, 'mfc')
 fh.make_new_directory(feature_dir, existing_dir='leave')
 tmp_dir = os.path.join(default.htk_dir, 'tmp')
@@ -65,11 +67,17 @@ fh.make_new_directory(tmp_dir, existing_dir='leave')
 label_dir = os.path.join(default.htk_dir, 'label')
 fh.make_new_directory(label_dir, existing_dir='leave')
 
+
+
 ## training
 hcompv_scp_train = os.path.join(tmp_dir, 'train.scp')
 mlf_file_train   = os.path.join(label_dir, 'train_phone.mlf')
 mlf_file_train_aligned = os.path.join(label_dir, 'train_phone_aligned.mlf')
 
+## testing
+htk_stimmen_dir = os.path.join(default.htk_dir, 'stimmen')
+
+
 ## train without sp
 niter_max = 10
 
@@ -100,7 +108,7 @@ if make_lexicon:
 
 
 ## intialize the instance for HTK.
-chtk = pyhtk.HTK(config_dir, fame_asr.phoneset_htk, lexicon_htk)
+chtk = pyhtk.HTK(config_dir, fame_asr.phoneset_htk, lexicon_htk, feature_size)
 
 
 ## ======================= make label files =======================
@@ -223,11 +231,14 @@ if extract_features:
 if flat_start:
 	timer_start = time.time()
 	print('==== flat start ====')
-	feature_size = 39
-	model0_dir	  = os.path.join(model_dir, 'hmm0')
 	fh.make_new_directory(model0_dir, existing_dir='leave')
 
-	chtk.flat_start(hcompv_scp_train, model0_dir, feature_size)
+	chtk.flat_start(hcompv_scp_train, model0_dir)
+
+	# create macros.
+	vFloors = os.path.join(model0_dir, 'vFloors')
+	if os.path.exists(vFloors):
+		chtk.create_macros(vFloors)
 
 	# allocate mean & variance to all phones in the phone list
 	print('>>> allocating mean & variance to all phones in the phone list...')
@@ -241,69 +252,38 @@ if flat_start:
 
 ## ======================= train model without short pause =======================
 if train_model_without_sp:
-	fh.make_new_directory(model1_dir)
-
 	print('==== train model without sp ====')
-	if not os.path.exists(os.path.join(model1_dir, 'iter0')):
-		shutil.copytree(model0_dir, os.path.join(model1_dir, 'iter0'))
-	for niter in range(1, niter_max):
-		timer_start = time.time()
-		hmm_n = 'iter' + str(niter)
-		hmm_n_pre = 'iter' + str(niter-1)
-		modeln_dir	   = os.path.join(model1_dir, hmm_n)
-		modeln_dir_pre = os.path.join(model1_dir, hmm_n_pre) 
-		
-		# re-estimation
-		fh.make_new_directory(modeln_dir)
-		pyhtk.re_estimation(
-			config_train,
-			os.path.join(modeln_dir_pre, hmmdefs_name), 
-			modeln_dir,
-			hcompv_scp_train, phonelist_txt,
-			mlf_file=mlf_file_train, 
-			macros=os.path.join(modeln_dir_pre, 'macros'))
-		print("elapsed time: {}".format(time.time() - timer_start))
+	
+	timer_start = time.time()
+	niter = chtk.re_estimation_until_saturated(
+		model1_dir,
+		model0_dir, improvement_threshold, hcompv_scp_train, 
+		os.path.join(htk_stimmen_dir, 'mfc'), 
+		'mfc', 
+		os.path.join(htk_stimmen_dir, 'word_lattice.ltc'), 
+		mlf_file=mlf_file_train, 
+		lexicon_file=os.path.join(htk_stimmen_dir, 'lexicon_recognition.dic')
+		)
+
+	print("elapsed time: {}".format(time.time() - timer_start))
 
 
 ## ======================= adding sp to the model =======================
 if add_sp:
 	print('==== adding sp to the model ====')
+	# reference:
+	# http://www.f.waseda.jp/yusukekondo/htk.html#flat_start_estimation
 
 	# make model with sp.
-	print('>>> modifying the last model in the previous step...')
-	modeln_dir_pre = os.path.join(model1_dir, 'iter'+str(niter_max-1))
-	modeln_dir = modeln_dir_pre.replace('iter' + str(niter_max-1), 'iter' + str(niter_max))
-	fh.make_new_directory(modeln_dir)
-	shutil.copy(
-		os.path.join(modeln_dir_pre, 'macros'), 
-		os.path.join(modeln_dir, 'macros'))
-	shutil.copy(
-		os.path.join(modeln_dir_pre, hmmdefs_name), 
-		os.path.join(modeln_dir, hmmdefs_name))
+	niter = 7
+	print('>>> adding sp state to the last model in the previous step...')
+	fh.make_new_directory(model1sp_dir, existing_dir='leave')
+	modeln_dir_pre = os.path.join(model1_dir, 'iter'+str(niter))
 
-	## ======================= 
-	## manually make changes to modeln_dir/hmmdefs 
-	## ======================= 
-	# add states 'sil'.
-	# http://www.f.waseda.jp/yusukekondo/htk.html#flat_start_estimation
-	#shutil.copy(
-	#	os.path.join(model_dir, 'hmmdefs.txt'), 
-	#	os.path.join(modeln_dir, hmmdefs_name))
-
-	#hmmdefs_file_pre = os.path.join(modeln_dir_pre, hmmdefs_name)
-	hmmdefs_file = os.path.join(modeln_dir, hmmdefs_name)
-	macros_file  = os.path.join(modeln_dir, 'macros')
-	#with open(hmmdefs_file_pre) as f:
-	#	lines = f.read()
-	#lines_ = lines.split('~h ')
-	#sil_model = [line for line in lines_ if line.split('\n')[0].replace('"', '') == 'sil'][0]
-	
-	# update hmmdefs and macros.
-	print('>>> updating hmmdefs and macros...')
-	modeln_dir_pre = modeln_dir
-	modeln_dir = modeln_dir.replace('iter' + str(niter_max), 'iter' + str(niter_max+1))
-	fh.make_new_directory(modeln_dir)
-	pyhtk.include_sil_in_hmmdefs(macros_file, hmmdefs_file, modeln_dir, sil_hed, phonelist_txt)
+	## update hmmdefs and macros.
+	print('>>> adding sp to the model...')
+	modeln_dir = os.path.join(model1sp_dir, 'iter0')
+	chtk.add_sp(modeln_dir_pre, modeln_dir)
 
 
 ## ======================= train model with short pause =======================