Changes between Initial Version and Version 1 of Ticket #41


Ignore:
Timestamp:
08/08/06 22:37:14 (14 years ago)
Author:
kmaclean
Comment:

Possible solutions:

  • use HTK 3.2
  • get mfc from Acoustic Model, and merge with adaptation data and create new Acoustic Model from scratch - would take too long ...

Legend:

Unmodified
Added
Removed
Modified
  • Ticket #41 – Description

    initial v1  
    66 
    77I have followed a similar approach to the Adaptation steps outlined in the following 2 email posts to HTK-Users, which talked about HERest supervised adaptation using HTK 3.3, but on a 'PLAIN' system (which I assume means monophones): 
     8 
     9{{{ 
    810 - http://htk.eng.cam.ac.uk/pipermail/htk-users/2006-July/005308.html 
    911 - http://htk.eng.cam.ac.uk/pipermail/htk-users/2006-July/005310.html 
    1012 
     13}}} 
     14 
    1115I am trying to adapt an Acoustic Model I created using the steps outline in the tutorial in the HTK 3.2 Book using HTK 3.3 and HERest. I don't think there is a problem with my data because I successfully adapted the same Acoustic Model (created using HTK3.3) using HTK 3.2 and HEAdapt. 
    1216 
    1317The approach I took was as follows: 
    1418 
    15 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ 
    16  Steps #1-11: Created Acoustic Models 
    17 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ 
     19 
     20 
     21==  Steps !#1-11: Created Acoustic Models == 
     22 
     23 
    1824Created Tied-State Triphones as per the tutorial in the HTK 3.2 Book 
    1925 
    20 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ 
    21  Step #12: Preparation of the Adaptation Data 
    22 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ 
     26 
     27 
     28==  Step !#12: Preparation of the Adaptation Data  == 
     29 
     30 
    2331(1) adaptWords.scp contains the list of adaptation 
    2432files 
     
    2937(3) config contains the basic configuration 
    3038variables for MFCC feature extraction: 
     39 
     40{{{ 
    3141 
    3242TARGETKIND = MFCC_0_D_N_Z 
     
    4050CEPLIFTER = 22 
    4151NUMCEPS = 12 
     52}}} 
     53 
    4254 
    4355(4) Realign the training data: 
    4456 
     57{{{ 
     58 
    4559HVite -A -D -V -T 1 -l '*' -o SWT -b SENT-END -C config -H hmm15/macros -H hmm15/hmmdefs -i adaptPhones.mlf -m -t 250.0 150.0 1000.0 -y lab -a -I adaptWords.mlf -S adapt.scp dict tiedlist  
    4660 
     61}}} 
     62 
    4763This creates adaptPhones.mlf 
    4864 
    49 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ 
    50  Step #13: Generating the Transforms  
    51 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ 
     65 
     66==  Step !#13: Generating the Transforms == 
     67  
     68 
    5269(1) regtree.hed as following: 
    5370  
     71 
     72{{{ 
    5473RN "models" 
    5574LS "stats" 
    56 RC 32 "rtree"  
     75RC 32 "rtree" 
     76}}} 
     77  
    5778  
    5879(2)  Generate Regression Class Tree: 
    5980 
     81 
     82{{{ 
    6083HHEd -A -D -V -T 1  -H hmm15/macros -H hmm15/hmmdefs -M hmmAdapt regtree.hed tiedlist 
     84}}} 
     85 
    6186 
    6287This creates 4 files in hmmAdapt: 
     
    6691        - rtree.base 
    6792 
    68 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ 
    69  Step #14: Global MLLR  
    70 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ 
     93 
     94==  Step !#14: Global MLLR == 
     95 
    7196(1) create a "glob" (under hmmAdapt folder) file 
    7297as follows:  
    7398 
     99 
     100{{{ 
    74101~b "glob" 
    75102<MMFIDMASK> * 
     
    77104<NUMCLASSES> 1 
    78105 <CLASS> 1 {*.state[2-4].mix[1]} 
     106}}} 
     107 
    79108 
    80109(2) config_adapt_global contains the configuration 
    81110variables for global transformation: 
    82111 
     112 
     113{{{ 
    83114HADAPT:TRANSKIND = MLLRMEAN 
    84115HADAPT:USEBIAS = TRUE 
     
    87118HADAPT:TRACE = 61 
    88119HMODEL:TRACE = 512 
     120}}} 
     121 
    89122 
    90123(7) tiedlist contains the list of HMMs 
     
    92125(8) Perform 1st pass Global Adaptation using HERest as following: 
    93126  
     127 
     128{{{ 
    94129HERest -A -D -V -T 1 -C config -C config_adapt_global -S adapt.scp -I adaptPhones.mlf -H hmmAdapt/macros -H hmmAdapt/hmmdefs -H hmmAdapt/glob -K hmmAdapt mllr -u a tiedlist 
    95130 
     131}}} 
     132 
    96133This creates a file called 'mfc.mllr' in hmmAdapt 
    97134 
    98 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ 
    99  Step #15: Adaptation using Regression Class Tree 
    100 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ 
     135==  Step !#15: Adaptation using Regression Class Tree == 
     136 
    101137(1) config_adapt_regc contains the configuration 
    102138variable for regression class tree transform: 
    103139 
     140 
     141{{{ 
    104142HADAPT:TRANSKIND = MLLRMEAN 
    105143HADAPT:USEBIAS = TRUE 
     
    109147HADAPT:TRACE = 61 
    110148HMODEL:TRACE = 512 
     149}}} 
     150 
    111151 
    112152(1) Perform 2nd pass global transformation: 
    113153 
     154 
     155{{{ 
    114156HERest -a -A -D -V -T 1 -C config -C config_adapt_regc -S adapt.scp -J hmmAdapt mllr -I adaptPhones.mlf -H hmmAdapt/macros -H hmmAdapt/hmmdefs -H hmmAdapt/mfc.mllr -K hmmAdapt mllr1 -u a -z regc.tmf tiedlist 
    115157 
     158}}} 
     159 
    116160When I run this I get the a segmentation fault as follows: 
    117161 
     162 
     163{{{ 
     164 
     165 
    118166HERest -a -A -D -V -T 1 -C config -C config_adapt_regc -S adapt.scp -J hmmAdapt mllr -I adaptPhones.mlf -H hmmAdapt/macros -H hmmAdapt/hmmdefs -H hmmAdapt/mfc.mllr -K hmmAdapt mllr1 -u a -z regc.tmf tiedlist 
     167 
    119168 
    120169HTK Configuration Parameters[17] 
     
    1652143 MMF input files 
    166215Segmentation fault 
     216}}} 
     217 
    167218 
    168219Any help would be greatly appreciated,