marytts-builder:
viewimport -> DatabaseImportMain -> viewImportComponent:(and child class)
lib: install by apt-get. praat package.
input: wav
output dir: data/pm
praat /home/research/data/tts/cmu/temp/script.praat /home/research/data/tts/cmu/basenames.lst /home/research/data/tts/cmu/wav/ /home/research/data/tts/cmu/pm/ 100 500
where script.praat is a script command.
praat is a lib for pitch marks
dst: generator pitchmarks, lpc, mel_cepstream
input: pm files
output dir: data/mcep/
hamming windows and so on
/usr/bin/sig2fv -window_type hamming -factor 2.5 -otype est_binary -coefs melcep -melcep_order 12 -fbank_order 24 -shift 0.01 -preemph 0.97 -pm /home/research/data/tts/cmu/pm/arctic_a0001.pm -o /home/research/data/tts/cmu/mcep/arctic_a0001.mcep /home/research/data/tts/cmu/wav/arctic_a0001.wav
data format transfer
AllophonesExtractor.java
request the server. it need to do a lot of nlp tagger.
dst: For the given texts, compute allophones, especially boundary tags.
output: prompt_allophones
set the request type and then call MaryHttpClient.java:_precess() --> requestInputStream() function to get the result then output to file.
using the allophones result to do labling. label the voice. they are all precessing the data/lab/ files.
process the label file and remove the continue 'pau'
dst: For the given texts, compute unit features and align them with the given unit labels.
It offer manully label tool which can select several token and its voice. If push the save button, it will save the correspond pm, wav, lab file in the temp folder.
It offer the method to play a selected section, or save the selection.
Convert phone labels to unit labels
output dir: data/phonelab
energy peak difference.. it will create the data/halfphonelab/
###transcriptionAligner
transcriptionAligner.java -> MaryTranscriptionAligner.java
dst: align and change automatic transcriptions to manually corrected ones
this will create the allophones folder. how to do the alignment, not very sure.
get the features-discrete and save to data/mary/features.txt
request from server. and the server precess it with InfoRequestHandler.java
extract context feature vectors from the text data. This procedure will create a phonefeatures directory
verify alignment between "phonefeatures" and "phonelabels".
timelinemaker for what?
HMMVoiceDataPrepare
only to prepare the data and set up the environment
parameter configs.
reads the mary/features.txt file (created in step 6), and generates the file mary/hmmFeatures.txt.
The hmmFeatures.txt file contains extra features, apart from phone and phonological features, that will be used to train HMMs. normally it is a subset of features.txt
call the hts to 'make data':make mgc lf0 str-mary cmp-mary list scp
HMMVoiceMakeData.makeLabels(): get label and mlf files. HMMVoiceMakeData.makeQuestions(): get questions files.
hts train.
log in the marytts/log. server run the marytts.runtime. it load a lot of module to process the text or voices. such as marytts-runtime/marytts.modules.OpenNLPPostTagger.java to do POS tagger.