ILive TTS Project

build-model.sh

#!/bin/sh
# ------------------------------------------------------------------------------
# FILE        | build-model.sh
# AUTHOR      | sameh kamal
# DESCRIPTION | build a new concatinative model
# ------------------------------------------------------------------------------

db_name = $1                # database name
dt_dir  = $2                # raw data dir
ws_dir  = "~/tts-workspace" # workspace dir
db_dir  = $ws_dir/$db_name  # database dir
aux_dir = $ws_dir/aux
phonemizer_continuous_dir = "~/univox/phonemizer-continuous"
marytts_base_dir          = "~/git/marytts"

# make workspace dir if not exists
mkdir -p $db_dir

if [ -d $db_dir ]; then
    # create database directories
    mkdir -p $db_dir
    mkdir -p $db_dir/text
    mkdir -p $db_dir/wav
    mkdir -p $db_dir/lab
    mkdir -p $db_dir/language

    # move raw data files to database directory
    cp $dt_dir/**/**/*.txt $db_dir/text
    cp $dt_dir/**/**/*.wav $db_dir/wav
    cp $dt_dir/**/*.lab    $db_dir/lab

    # enter database directory
    cd $db_dir
    # -------------------------------------------------------------------
    # create unique words file
    # -------------------------------------------------------------------
    # enter text directory
    cd text
    # put all sentences in one file in language directory
    awk 'FNR==1{print ""}1' *.txt > ../language/text_ar.txt
    # enter language directory
    cd ../language
    # get unique words from all sentences and put them in one file
    tr -s [:space:] \\n < text_ar.txt | sort | uniq > unique_words.txt
    # remove space and ',' from the head of file if exists
    ln     = $(head -n 1 unique_words.txt)
    ln_len = ${#ln}
    if [ $ln_len eq 1 ]
       sed -i '1d' unique_words.txt
    fi
    ln     = $(head -n 1 unique_words.txt)
    ln_len = ${#ln}
    if [ $ln_len eq 1 ]
       sed -i '1d' unique_words.txt
    fi
    # -------------------------------------------------------------------
    # replace SSIL and SIL with _ in labs files   
    # -------------------------------------------------------------------
    cd ../lab
    # replace all SSIL with _ in all files
    sed -i -- 's/SSIL/_/g' *
    # replace all SIL with _ in all files
    sed -i -- 's/SIL/_/g' *

    # -------------------------------------------------------------------
    # run phonemizer on unique words   
    # -------------------------------------------------------------------
    cd ../language
    java -jar $aux_dir/phonemize_cont.jar ./unique_words.txt ./ar_phon_dict.ph
else
    echo "> database dir doesn't exist. "
fi

ilive-tts.md

iLive TTS model generation

This is a bief explanation of how to build a new model for ilive tts.

input

the input of model generation process will be a set of directories containing required files for model generation, and these folders can be listed as the follwing:

• text dir contains set of diactized arabic text sentences each in a separated file.
• wav   dir contains set of wav files each represents arabic pronounciation of corresponding sentence in text dir
• lab dir contains arabic text pronounciation timelaps for each speech segment for each corresponding text and wav file
• language dir we create that contain pre-processing outputs
• phonemizer dir contain set of rules of pronounciation of arabic language phonemes.

Pre-Procesing

the input acquired from linguistics team needs some pre-processing to generate intermediate files needed while generating the model.

• gather all unique words in all scentences in one file and put it in language folder.

# enter text directory
cd text
# put all sentences in one file in language directory
awk 'FNR==1{print ""}1' *.txt > ../language/text_ar.txt
# enter language directory
cd ../language
# get unique words from all sentences and put them in one file
tr -s [:space:] \\n < text_ar.txt | sort | uniq > unique_words.txt
# then manually remove space` ` and `,` if exist in the unique words file

• in lab directory, in all files, replace all SIL and SSIL with _.

# enter lab directory
cd lab
# replace all SSIL with _ in all files
sed -i -- 's/SSIL/_/g' *
# replace all SIL with _ in all files
sed -i -- 's/SIL/_/g' *

• in project phonemizer-continuous resources [src/main/resources/com/univox], replace [allophones.ar_SA.xml, ArabicPhonemesMap, ArabicScript] with files from phonemizer directory, and also in both the jar file and the rar file inside it.
• in project marytts:marytts-runtime/src/main/resources and marytts-runtime/src/main/java/marytts/com/univox replace [ArabicPhonemesMap,ArabicScript,allophones.ar_SA.xml] with files from the previos step.
• put the unique words file in project phonemizer-continuous base folder.
• in project phonemizer-continuous:src/main/java/com/univox/PhonemizerMain.java make sure that the file name used iis the same as your unique words file name in the base filder.

String filename="unique_words.txt";

• run project phonemizer-continuous as a java application. [this will take a few minutes]
• move the output of phonemizer-continuous project output named unique_words.ph to language directory and name it ar.txt.
• in language folder, in ar.txt file replace all __ with functional (yes: with a pre space), then remove all the remaining _ from the file.

# enter language folder
cd language
# replace "__" with " functional"
sed -i -- 's/__/ functional/g' ar.txt
# remove remaining '_' in the file
sed 's/_//g' ar.txt

• in marytts project, delete target folder, then re-create marytts

# enter marytts directory
cd marytts
mvn -Dmaven.test.skip=true install

• in language folder run transcription.sh which results from marytts build in the previous step.

cd language
transcription.sh

this will open a GUI tool that will require few steps:-

• asks for alophones.ar_SA.xml file, select it from its location.
• then from file menu select open and then select ar.txt file from language directory.
• check all words, none should be in red, and if so this indicates and error.
• click train and predict button.
• from file menu, select save. this will result in saving few files in language directory.
• close the gui tool.
• in marytts project, delete target folder.
• in marytts project, in directory marytts-languages/marytts-lang-ar, delete target folder.
• in marytts project, in directory marytts-languages/marytts-lang-ar/src/main/resources/marytts/language/ar/lexicon replace the files in it with output files from the transcription.sh tool.
• in the directory from the previous step rename allophones.ar_SA.xml to allophones.ar.xml and also remove the _SA from the tag lang inside the file.
• in project marytts: dir marytts-language/marytts-lang-ar/lib/modules/ar/lexicon repace the two files [allophones.ar.xml,ar] with modified all_phones.ar.xml file from the previos step and ar.txt file from language directory afrer being renamed to ar only
• in project marytts: dir marytts-language/marytts-lang-ar test that everything is okay.

mvn test

• in project marytts:dir user-dictionaries, replace userdict-ar.txt with the output of the phonemizer-continuous project which is unique_words.ph after it being renamed to userdict-ar.txt.
• edit the filed moved to user-dictionaries, replace remove __ and replace _ with |

sed 's/__//g' userdict-ar.txt
sed 's/_/| /g' userdict-ar.txt

• in marytts project, delete target folder, then build marytts

# enter marytts directory
cd marytts
mvn -Dmaven.test.skip=true install

• open marytts projcet - univox workspace -in eclipse and add build configuration for server and client as the following.
• run server: run configuration with ~/git/marytts as mary base
• run DatabaseImportMain: run configuration with db directory as database base folder.

set-env-vals.sh

#!/bin/sh
# ------------------------------------------------------------------------------
# FILE        | set-env-vals.sh
# AUTHOR      | sameh kamal
# DESCRIPTION | set environment variables required for building tts models
# ------------------------------------------------------------------------------

export MARY_BASE_DIR=/home/sameh/work/ist
export PHONEMIZER_SCRIPTS_DIR=/home/sameh/work/ist/phonemizer