Skip to content

Latest commit

 

History

History

am-train

README.md

HTK - tutorial para treino de modelo acústico

For this tutorial we are following the chapter 3 of the HTKBook for HTK version 3.4. Installation on Arch linux required lib32-glibc and lib32-gcc-libs packages as dependencies.

graph LR;
  A[data<br>preparation]-->B[creating<br>monophone HMMs];
  B[creating<br>monophone HMMs]-->C[creating<br>tied-state<br>triphone HMMs];
  C[creating<br>tied-state<br>triphone HMMs]-->D[recogniser<br>evaluation];
Loading

3.0 FalaBrasil Environment Tree

HLEd                                                HTK Label Editor   
  -l '*'                                            dir to store label file (MLF header)   
  -d etc/dictionary.dic                             load phonetic dict   
  -i etc/phones0.mlf (or phones_sp.mlf)             output MLF transcription to file    
  $MYHTKPROJ/util/mkphones.led (or mkphones_sp.led) command editor file    
  $MYHTKPROJ/etc/words.mlf                          label file

HCopy                                "copy" an in file to an encoded out file
  -T 1                               set trace flag
  -C conf/edaz.conf                  set config file
  $MYHTKPROJ/wav/train/file.wav      input audio file
  $MYHTKPROJ/wav/train/file.mfc      output parameterized audio file
3.1.4: Step 4 - Creating the Transcription Files 3.1.5: Step 5 - Coding the Data

3.2 Creating Monophone HMMs (fb_02_create_monoph_hmm.sh)

3.2.1 Step 6 - Creating Flat Start Monophones

HCompV
  -C conf/hcomp.conf
  -f 0.01
  -m
  -S etc/train\_mfc.scp 
  -M model/flat\_start 
   model/flat\_start/proto
3.2.1: Step 6 - Creating Flat Start Monophones 
                  out                     in dir             in                 in

HHEd -w model/flat_start/hmms.mlf -d model/flat_start util/concatenade.hed etc/hmmlist.txt

        in                                in                  in                           in                            out dir        in

HERest -I etc/phones0.mlf -t $PRUNING -S etc/train_mfc.scp -H model/flat_start/macros -H model/flat_start/hmmdefs -M model/mono_0 etc/hmmlist.txt

3.2.2 Step 7 - Fixing the Silence Models

              in                         in                         out_dir              in           in

HHEd -H /model/mono_sp_0/macros -H /model/mono_sp_0/hmmdefs -M /model/mono_sp_1 /util/sil.hed /etc/hmm_sp.list

                        in                        in                     in                                   in                            out_dir             in

HERest -t $PRUNING -I /etc/phones_sp.mlf -S /etc/train_mfc.scp -H /model/mono_sp_$((i-1))/macros -H /model/mono_sp_$((i-1))/hmmdefs -M /model/mono_sp_${i} /etc/hmm_sp.list

3.2.3 Step 8 - Realigning the Training Data

                                           in                    out                         in                                in                           in                   in              in

HVite -o SWT -b sil -a -y lab -m -t 250.0 -I /etc/words.mlf -i /etc/aligned.mlf -H /model/mono_realign_0/macros -H /model/mono_realign_0/hmmdefs -S /etc/train_mfc.scp /etc/dictionary.dic /etc/hmm_sp.list

                      in                  in                          in                                     in                                      out_dir                in

HERest -t $PRUNING -I etc/aligned.mlf -S etc/train_mfc.scp -H model/mono_realign_$((i-1))/macros -H model/mono_realign_$((i-1))/hmmdefs -M model/mono_realign_${i} etc/hmm_sp.list

3.3 Creating Tied-State Triphones

FalaBrasil UFPA

Grupo FalaBrasil (2019) - https://ufpafalabrasil.gitlab.io/
Universidade Federal do Pará (UFPA) - https://portal.ufpa.br/
Cassio Batista - https://cassota.gitlab.io