-------------------------------------------------------- LIA_PHON : Un systeme complet de phonetisation de textes -------------------------------------------------------- Copyright (C) 2001 FREDERIC BECHET .................................................................. This file is part of LIA_PHON LIA_PHON is free software; you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation; either version 2 of the License, or (at your option) any later version. This program is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details. You should have received a copy of the GNU General Public License along with this program; if not, write to the Free Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA .................................................................. For any publication related to scientific work using LIA_PHON, the following reference paper must be mentioned in the bibliography: Bechet F., 2001, "LIA_PHON - Un systeme complet de phonetisation de textes", revue Traitement Automatique des Langues (T.A.L.) volume 42, numero 1/2001, edition Hermes .................................................................. Contact : FREDERIC BECHET - LIA - UNIVERSITE D'AVIGNON AGROPARC BP1228 84911 AVIGNON CEDEX 09 FRANCE frederic.bechet@lia.univ-avignon.fr .................................................................. ------ README ------ I : Content of the package LIA_PHON II : Installation III : Useful scripts IV : Quick overview of each LIA_PHON command V : Data format VI : Warning I - Content of the package LIA_PHON ----------------------------------- Once you have uncompress and untar the file 'lia_phon.tar.gz', a directory 'lia_phon' is created, containing the following files and directories: - Makefile : compiling sources and data files of LIA_PHON - README : this present file - /bin : directory which will contain the executable files - /data : directory containing the following data files - desigle.pron : decision spelled/read for acronyms - epeler_sig.pron : pronunciation rules for spelled acronyms - lire_sig.pron : pronunciation rules for read acronyms - french01.pron : pronunciation rules for standard French - h_aspi.sirlex : exception list of words starting with an 'h' - initfile.lia : correspondence phonetic alphabets MBROLA/LIA_PHON - lex10k : 10000 words lexicon used by the POS tagger - lex80k ; a 80k word lexicon used by the POS tagger - list_chif_virgule : list of words possibly following a digit string - list_exep : pronunciation exceptions of 'french01.pron' rule database - lm3classe.arpa : ARPA-standard LM for POS tagging produced by CMU-Cambridge SLMT - model_[un,bi,tri] : 3-letter models for the proper-name identification - model_morpho.[un,bi,tri] : 3-letter models for out-of-vocabulary words in POS tagging - propername_[1-8].pron : pronunciation rules for different proper-name sets - regles_l.pro3 : rules managing the liaison insertion between words - regles_retik : post-processing POS tagging specific to Text-to-Speech transcription - rule_phon.pro : post-processing rules for the phonetic strings - rule_variante.pro : rules allowing to generate alternate pronunciations - /doc : documentation about LIA_PHON as follows - gnu_gpl.txt : GNU General Public License Version 2 - lia_phon.htm : complete description of LIA_PHON components - test.txt : test file - test.ola : result you should obtain from the test file if everything is OK - /object : object files generated - /script : script command for using LIA_PHON - lia_make_lexique_reacc : compilation of the lexicon data of the POS tagger - lia_nett : tokenize in word and sentence, process digit string - lia_delete_lexicon : delete all the compiled files of a given lexicon - lia_taggreac : POS tagging + text accentuation - lia_phon : grapheme to phoneme transcription - lia_text2phon : text to phonetization (LIA_PHON format) - lia_text2mbrola : text to speech with prosodic information (MBROLA format) - lia_lex2phon : phonetization of a lexicon for all the possible POSs of each word - lia_genere_phrase_mbrola : one MBROLA file per sentence - /src : source files contained in the following directories - aphon : grapheme to phoneme transcription - format : tokenization - liaison : liaison generation - libgram : n-gram LM - post_process : phonetic string post-processing - prosodie : prosodic information generation - retik : specific tagging to grapheme to phoneme transcription - tagg : POS tagging + text accentuation - variante : alternate pronunciation generation II - Installation ----------------- LIA_PHON is composed of a set of modules all written in standard C, using only standard libraries. It has been successfully compiled on the following UNIX environments: - HP-UX - SUN-SOLARIS - IRIX - LINUX 1) To install the package, you must first set the environment variable: 'LIA_PHON_REP' with the path of the root directory 'lia_phon'. For example, in c-shell: setenv LIA_PHON_REP /usr/tools/lia_phon 2) Then edit the file 'Makefile' in order to change, if necessary the C-compiler and the compilation options which will be used to compile the package. 3) Compile the package with the command: 'make all' 4) Compile the resources with the command: 'make ressource' 5) Check is everything is OK with the command 'make check' 6) If you want to use the 80k word lexicon instead of the default 10k word lexicon (unnecessary for text-to-speech application but useful for the phonetization of a lexicon) just add the following commands: make lex80k setenv LIA_PHON_LEX lex80k That's all !! If you want to suppress the compiled files, just execute: make clean make clean_ressource III - Useful scripts -------------------- - Text-to-speech scripts LIA_PHON contains a set of scripts that transform a raw text into its phonetic form, which can be used as an input for the speech synthesizer MBROLA. There are four main steps in this process: 1) Formatting the text 2) POS tagging + accentuation 3) Grapheme-to-Phoneme transcription 4) MBROLA format Three scripts summarize all these steps: 1- lia_text2phon input = stdin output= stdout Take as input a raw text and output a tagged, phonetized text in the LIA_PHON format For example: $LIA_PHON_REP/script/lia_text2phon < test.txt > test.pho 2- lia_text2mbrola input = stdin output= stdout Take as input a raw text and output a MBROLA file with very very basic prosodic information. (see http://tcts.fpms.ac.be/synthesis/ for information about MBROLA speech synthesizer) For example: $LIA_PHON_REP/script/lia_text2mbrola < test.txt > test.ola Then, to produce a '.wav' file with MBROLA, one can use: $MBROLA_REP/mbrola -I $LIA_PHON_REP/data/initfile.lia $MBROLA_REP/fr1 test.ola test.wav 3- lia_genere_phrase_mbrola input = filename output = filename.i.ola Take as input a text filename and output N '.ola' files for each sentence of the text file. All these scripts accept the option '-reacc'. If this option is set, the text-to-speech system considers that some accents might be missing on some words and tries to correct them. However this can lead to some accentuation errors. If you are sure that your text has all its accent correct, it's better not to use this option. - Phonetization of a lexicon LIA_PHON contains also a script which takes as input a list of words and output one or several phonetization for each word. The input format is a text file with one word on each line. The output format is a text file with, on each line, a triplet: <[POS]> If a word has several POSs in the LIA_PHON_LEX lexicon, one phonetization will be proposed for each POS, even if the phonetization are similar. For example, if the input file contains: mange couvent The output file will be: mange mmanjj [V2S] mange mmanjj [V3S] mange mmanjj [V1S] couvent kkouvv [V3P] couvent kkouvvan [NMS] For the acronyms and the proper names, only one phonetization is proposed. The command which performs this process is: $LIA_PHON_REP/script/lia_lex2phon where the input file is given on the standard input and the result in the standard output. IV - Quick overview of each LIA_PHON command -------------------------------------------- Each step in the text processing will be now illustrated on this small raw text example: -- Ceci est un test. De plus en plus, le 1 janvier tombe un lundi. Il couttent vrraiment 1,08 euros. Il eleve un debat deja eleve sur le TALN. -- 1) Formatting the text This is done with the script: $LIA_PHON_REP/script/lia_nett - input = stdin - output= stdout This script tokenize the text according to the lexicon used by the POS tagger. Then, the text is split into sentences, marked by the tags (beginning of a sentence) and (end of a sentence). The eventual capital letter of each first word of a sentence is then removed. Finally, digit strings are converted into words with respect to the symbol ',' and the output text contains one word on each line. This is the result of this script on the previous text: ceci est un test . de_plus_en_plus , le premier janvier tombe un lundi . il couttent vrraiment un virgule zéro huit euros . il eleve un debat deja eleve sur le TALN . 2) POS tagging + accentuation This is done by the script: $LIA_PHON_REP/script/lia_taggreac - input : stdin - output : stdout This script takes the output of 'lia_nett' and process a POS tagging and a text accentuation. The out-of-vocabulary words are processed by a POS guesser (for example, the words 'couttent' and 'vrraiment' in the previous example), except for unknown proper-names (capitalize words within a sentence) which remain with the tag 'MOTINC'. This is the result of this script on the output of 'lia_nett': ZTRM ceci PDEMMS est VE3S un DETMS test NMS . YPFOR ZTRM ZTRM de_plus_en_plus ADV , YPFAI le DETMS premier AMS janvier NMS tombe V3S un DETMS lundi NMS . YPFOR ZTRM ZTRM il PPER3MS couttent V3S vrraiment ADV un DETMS virgule CHIF zéro CHIF huit CHIF euros NMP . YPFOR ZTRM ZTRM il PPER3MS élève V3S un DETMS débat NMS déjà ADV élevé AMS sur PREP le DETMS TALN MOTINC . YPFOR ZTRM 3) Grapheme-to-Phoneme transcription This is done with the script $LIA_PHON_REP/script/lia_phon - input : stdin - output : stdout This script takes as input the output of 'lia_taggreac' and phonetize, insert liaison between words, decide to read or spell acronyms, calculate a pronunciation for proper-names, and check for exceptions to the transcription rules. This is the result of this script on the output of 'lia_taggreac' on the previous text: ## [ZTRM->EXCEPTION] ceci sseessii [PDEMMS] est ei [VE3S] un ttun [DETMS] test ttaissttee [NMS] pause ## ## [ZTRM->EXCEPTION] ## [ZTRM->EXCEPTION] de_plus_en_plus ddeepplluuzzanpplluuss [ADV] pause ## le llee [DETMS] premier pprreemmyyei [AMS] janvier jjanvvyyei [NMS] tombe ttonbb [V3S] un un [DETMS] lundi llunddii [NMS] pause ## ## [ZTRM->EXCEPTION] ## [ZTRM->EXCEPTION] il iill [PPER3MS] couttent kkoutt [V3S] vrraiment vvrrrraimman [ADV] un un [DETMS] virgule vviirrgguull [CHIF] zéro zzeirrau [CHIF] huit uyii [CHIF] euros tteurrau [NMP] pause ## ## [ZTRM->EXCEPTION] ## [ZTRM->EXCEPTION] il iill [PPER3MS] élève eillaivv [V3S] un un [DETMS] débat ddeibbaa [NMS] déjà ddeijjaa [ADV] élevé eilleevvei [AMS] sur ssuurr [PREP] le llee [DETMS] TALN ttaallnn [MOTINC->SIGLE->SIGLE_LU] pause ## ## [ZTRM->EXCEPTION] ???? [] 4) MBROLA format In order to use the MBROLA speech synthesizer (see http://tcts.fpms.ac.be/synthesis/ for more details), the output of 'lia_phon' can be formatted into MBROLA format, by adding some (minimal) prosodic information to the phonetic output. WARNING: this prosodic generation is here just to avoid listening to a 'flat' voice, but have no pretension to reflect a 'realistic' prosody !! To format the output of 'lia_phon' you must use the command: $LIA_PHON_REP/bin/lia_add_proso - input : stdin - output : stdout V - Data format --------------- LIA_PHON uses several kind of resources: 1- Grapheme-to-Phoneme transcription rule databases 2- Lexicons 3- Part-Of-Speech resources 4- Tagging and post-processing rules You can update any of these resources and the following quick description of each of them should answer most of the questions. BUT, don't forget to compile again the resources each time you modify one of the data files. 1- Grapheme-to-Phoneme transcription rule databases The main database is called 'french01.pron' and contains more than 1000 rules for the pronunciation of standard French. New rules can be added anywhere as the database is sorted each time it is loaded. The format of the rules of 'french01.pron' is as follow: regle(num,,pho,cont,ex) -> ; num : identification number of the rule left : left context of the grapheme to process graf : grapheme to transcribe cut into syllable right : right context of the grapheme to process pho : phonetic string corresponding to the grapheme to process cont : morphological constraint (verb V or non-verb NV) ex : example The following symbols can be used in the graphical context : V : any vowel C : any consonant S : 's' or nothing The grapheme are encoded as ASCII ISO8859-1 WARNING: the grapheme strings MUST be cut into syllables, for example, the word 'balbutier' will be split into 'bal bu tier' The phonetic alphabet used is describe as follows (with the correspondence with the SAMPA alphabet used in the MBROLA system) : LIA SAMPA EXAMPLES ii i idiot, ami ei e ému, été ai E perdu, maison aa a alarme, patte oo O obstacle, corps au o auditeur, beau ou u coupable, loup uu y punir, élu EU 2 creuser, deux oe 9 malheureux, peur eu @ petite, fortement in e~ peinture, matin an a~ vantardise, temps on o~ rondeur, bon un 9~ lundi, brun yy j piétiner, choyer ww w quoi, fouine pp p patte, repas, cap tt t tête, net kk k carte, écaille, bec bb b bête, habile, robe dd d dire, rondeur, chaud gg g gauche, égal, bague ff f feu, affiche, chef ss s soeur, assez, passe ch S chanter, machine, poche vv v vent, inventer, rêve zz z zéro, raisonner, rose jj Z jardin, manger, piège ll l long, élire, bal rr R rond, charriot, sentir mm m madame, aimer, pomme nn n nous, punir, bonne ## _ (silence marker) Several other database are used by LIA_PHON, with the same format: - desigle.pron : decision spelled/read for acronyms - epeler_sig.pron : pronunciation rules for spelled acronyms - lire_sig.pron : pronunciation rules for read acronyms - french01.pron : pronunciation rules for standard French - propername_[1-8].pron : pronunciation rules for different proper-name sets 2- Lexicons There is 3 lexicons used by LIA_PHON (except the POS resources): - h_aspi.sirlex Exception list of words starting with an 'h' Format of each line: indice\tword - list_exep Pronunciation exceptions of 'french01.pron' database Format of each line: word phonetic_string - list_chif_virgule List of words possibly following a digit string Format of each line: word 3- Part-Of-Speech resources The main editable resource of the POS tagging is the lexicon. The default one given with LIA_PHON contains 10K words, and is stored in the text file: lex10K The format of the dictionary is as follows: - one word on each line - format of each line: word POS1 freq(word,POS1) lemma(word,{POS1) POS2 freq(word,POS2) lemma(word,POS2) .... For example: assise AFS 213 assis NFS 733 assise VPPFS 320 asseoir To add a new word, just add a new line at this file, and if you don't have a corpus for estimating the frequency of each couple (word,POS), just put '1' instead. There's also a 80k word lexicon which can be used instead of the default one (not necessary for standard text-to-speech application). To use it, you have to perform the folllowing commands (in the LIA_PHON_REP directory): make lex80k setenv LIA_PHON_LEX lex80k 4- Tagging and post-processing rules These rules post-process the phonetic strings produced by LIA_PHON, calculate a possible liaison between two words and add new tags specific to the Grapheme-to-Phoneme transcription process. They are stored in the files: rule_phon.pro, regles_retik and regles_l.pro3 VI - Warning Be careful of the accent encoding of the files !! Check that the accents of the rule file 'french01.pron' are correct (ASCII ISO8859-1). If not, you have to use an accent transcoder to correct them. This transcoding MUST be applied to ALL the files in the directories /data and /src Contact ------- If you have any problem or question, contact me at : frederic.bechet@lia.univ-avignon.fr