Getting from JANUS-2 to JANUS-3

0. Evaluate and Copy your V2 System

• Consider, at least for a moment, not to convert your existing system but rather train up a completely new system.
• Make the JANUS-2 ystem you're going to convert use TRadd_dist 2 and TRtopN 0. This is the only score computation supported by v3, thus if you use it, comparisions will be easier. Also TRcbfact 1.0 and TRdsfact 1.0 are advisable.
• Run recognition with the JANUS-2 system you want to convert on your default testset.
  Make sure that you get verbose hypothesis output to stdout. Also, make sure to pipe the output of the recognizer to a file. Measure the Word Accuracy of this system. Don't touch this directory again until you're done converting.
• Next, make a complete copy of your V2 system so you don't destroy it. From now on, always work with the copy.
• If in doubt about file formats, don't trust this document but refer to the file format description of the appropriate module.

1. File conversion - Part 1: compatible Files

• Dictionaries: no changes
  
• Vocabularies: no changes
  
• Codebook Data & Description files: no changes
  
• Distribution Data & Description files: no changes
  
• LDA-Matrix: no changes
  
• Language Model Files:
  The format remains the same as for JANUS-2, exept that JANUS-3 also accepts compressed and gnuziped files. Note that if the filename is given as xxx, JANUS-3 will FIRST look for BINDUMP versions in the following order:
  • xxx.BINDUMP
  • xxx.BINDUMP.Z
  • xxx.BINDUMP.gz
  • xxx
  • xxx.Z
  • xxx.gz
  If a matching file is found, all further down the list are ignored. So it's no good to edit xxx unless you throw away xxx.BINDUMP. BINDUMP files are created from normal nist language model files automatically, whenever a normal nist file is read. BINDUMP files can do everything a normal nist file can do -- exept that you can't edit them. They load much faster than ordinary files.

2. File conversion - Part 2: simple Cases

• Senone Description File:
  If you have a one stream system, you don't have to change this file. If you have a multi-stream system, you have put braces arount the senon-names. This is because the info for all streams is one parameter for a tcl command that adds the info to an internal structure. Example:

      JANUS-2:
       #QK-m(1) #QK-m(1)-0 #QK-m(1)-1
        
      JANUS-3:
       #QK-m(1) {#QK-m(1)-0 #QK-m(1)-1}
     

• Transition File
  This format too now looks a little tcl-like. On each line of the input file, there is first the name of the transition model, and then one parameter that specifies what it does. This parameter is is a tcl-list, hence it goes into braces. Each item in the braces specifies stepwidth and penalty for one transition. The number of transitions is no longer required.
  Example:
  

      JANUS-2: 
       SIL           2
                     0  0.01
                     1  0.0
  
       3             3
                     0  0.01
                     1  0.0
                     2  0.015
  
      JANUS-3: 
       SIL  { { 0 0.01 } { 1 0.0  } }
       3    { { 0 0.01 } { 1 0.0  } { 2 0.015 } }
     

• Feature Description File:
  The feature description file now really is a tcl script. It still looks pretty much like the old feature description files. As most systems use pretty much the same preprocessing, here is an example of an old JANUS-2 converted to a JANUS-3 file. You should be able to cut and paste this to your needs. To find out more about the JANUS-3 feature module and methods, read the appropriate part of this document. Cutting off the first and last two frames of each file shows the flexibility of the new concept and is only done to provide full compatibility to the JANUS-2 system, where these frames are ignored to avoid problems with the coefficient involving delta and delta delta computation. Try the FeatureSet method "fs eval" and "fs show" to see if everything comes out allright.

     JANUS-2
      ; Feature Description File for JANUS-2 Training and Testing
      ;
      ; to      magic   what        from        parameters
      ;
      INPUT      21  2  read        FFT
      MFSC       16  3  cut         INPUT       0 15        
      dMFSC      16  4  delta       MFSC        2
      LOGPOW      1 11  cut         INPUT       16 16
      dLOGPOW1    1  5  delta       LOGPOW      1
      ddLOGPOW1   1  6  delta       dLOGPOW1    1
      dLOGPOW2    1  7  delta       LOGPOW      2
      ddLOGPOW2   1  8  delta       dLOGPOW2    2
      COMBI       4  9  cut         INPUT       17 20
      ZWISCH     40 10  concat      MFSC COMBI
      LDAS       40 13  meanvarnorm ZWISCH      2
      LDAT       40 12  transform   LDAS ../env-GSST/matrix.cd
      LDA1       16  0  cut         LDAT        0 15       
      LDA2       16  1  cut         LDAT        16 31       
     
     JANUS-3
      # Feature Description File for JANUS-3
      #
      #    what     to        from           parameters
      #
      $fes read    INPUT      $arg(FFT)
      $fes split   MFSC       INPUT           0 15
      $fes delta   dMFSC      MFSC            -delta -2
      $fes split   LOGPOW     INPUT           16 16
      $fes delta   dLOGPOW1   LOGPOW          -delta -1
      $fes delta   ddLOGPOW1  dLOGPOW1        -delta -1
      $fes delta   dLOGPOW2   LOGPOW          -delta -2
      $fes delta   ddLOGPOW2  dLOGPOW2        -delta -2
      $fes split   COMBI      INPUT           17 20
      $fes merge   ZWISCH     MFSC dMFSC dLOGPOW1 ddLOGPOW1 dLOGPOW2 ddLOGPOW2 COMBI
      $fes meansub LDAS       ZWISCH          -a 2
      $fes matmul  LDAT       LDAS            $fes:matrixfeat.data
      $fes split   LDA1-big   LDAT            0 15
      $fes split   LDA2-big   LDAT            16 31
      set  frameN [expr [$fes:LDA1-big configure -frameN] -3]
      $fes cut     LDA1       LDA1-big        2 $frameN 
      $fes cut     LDA2       LDA2-big        2 $frameN 
     

3. File conversion - Part 3: new Files

• Tags:
  Those are not used much yet. However, you better create a single file called something like "tags" and put two lines into it:

      WB
      WE
    

• Topology Description
  In JANUS-3, the way states can be arranged in a phoneme is no longer hardcoded. Therefore, we have a set of different models that can later be refered to with their names. The corresponding file for a model called 3state and a model called SIL looks like this:

      3state { ROOT-b ROOT-m ROOT-e } { 3 3 3 }
      SIL    { ROOT-m ROOT-m }        { SIL SIL }
    

  The SIL and the 3 that are the last pair of braces refer to the names in the transition file. Together with the transition file this defines a set of full phoneme topologies.

• Topology Tree
  Now we need to decide which topology model to use for which phoneme in which context. This is aceived by a topology tree. Each line in this file has 6 fields: The node name, the question, the no-sucessor, the yes-sucessor, the dont-know-successor and the answer. If the question is empty the line represents a tree leaf. For reasons we won't go into now, tree nodes can have answers too. Also you don't have to worry about the dont-know-successor yet. For my standard example, the topology tree (or ttree) would look like this:

     ttree      { 0=SIL } 3stateNode silNode  -   -
     3stateNode {       }    -         -      -   3state
     silNode    {       }    -         -      -  SIL
    

  Note that the questions come in braces. This is because you can put a list of qustions that are connected by and. The 0 in the first line means the question refers to the current phoneme, not the one left or right of it. 0=SIL means, is the current phoneme a SIL Phoneme ? If no, we go to the leaf that answers to use the model named 3state for this phoneme, if yes we go to the leaf with the model named SIL.

• Data Bases
  You don't need databases to use JANUS-3, but you're strongly encouraged to use them. They take the place of the utterance files in JANUS-2 and allow faster access to large amount of data. You best read the appropriate documentation.

4. File conversion - Part 4: the tricky bits

• Questions File and Phoneset
  The question file is no longer used in JANUS-3. The information in it goes partly into the new phonset file, and into the new senon trees. Make sure to keep a copy of the v2 questions file around -- it's required for converting the tree file. You can create a phonset file manually, but if your questions contain a lot of defines and use a lot of classes on the right hand side, you might want to try using the newtree script for that too (see below).

• The senon Tree
  As phoneme topologies are handled elsewhere in JANUS-3, you need a senon tree. To build one, you need a system that would start up but for the senon tree and the phoneset file. The easiest way is to steal a setup from another task. There is an example directory for JANUS-3 with GSST currently in Monikas home directory, that will sometime soon move into the main janus-3 path. It contains scripts that work well for GSST, but should also do for similar task. If you've never worked with JANUS-3, you might want to be able to see how it works before you continue, or find a JANUS-3 GURU to help you now.

  Ready ? Ok.

  • Make sure you know where the old tree and questions file is.
  • Start JANUS-3 interactively.
  • source the script newtree.tcl (from my example diretory)
  • Build a phonemeSet file by calling:
    basetree tree questions badtree phoneset
    where tree is the full filename of the JANUS-2 tree and questions it the full filename of your JANUS-2 questions. As a result you get the files "badtree" and "phoneset" If there are any problems with this step, there is little use to continue without help.
  • The file "phoneset" should not contain any 'WB' and also should contain the class PHONES and the class any, each representing all monophones. Edit the file to remove WB's and add those classes.
  • Create a testscript that will initialize a complete v3 system. You can start building one with the testschedule provided in the example system. Note that it sources other files from the script directory too, so you better get all scripts in it. Use 'badtree' and 'phoneset' as files for the tree and the phoneset.
  • Make sure you have about 3 Meg diskspace available. The new tree will be huge.
  • Start JANUS-3 interactively again, source newtree.tcl and then source your testscript. Don't continue until this goes through without and error message.
  • Note the numbers behind RT.nodeN after creating the search object. It's probably NOT the same as you had in your last JANUS-2 experiment. This is because your tree is still all wrong (though it has the correct format for JANUS-3 to read).
  • Run:
    treeDoall stree0 tree questions goodtree deleteme
    Where stree0 is the name of your current senon tree object, tree is the filename of your v2 tree, questions is the filename of your v2 questions, goodtree is the output of a good tree, and deleteme should be deleted after this. This will take a while to complete. Errors of the type
    Can't correct ? ANG SIL
    can be ignored if they always contain a SIL phoneme. If there are any other problems, try to run the steps in the routine treeDoall manually one by one.
  • Restart JANUS-3 this time using the "goodtree". See if the value behind RT.nodeN is correct now. If it's not, you're in trouble. Find a Guru.
  • Run a test on your favorite testset. Results should be the same as for your JANUS-2 Version. Scores might differ in the order of magnitude of 0.1%.