The twisted Tongue
Contents
1. fuge nill is the empty string Consider a first approximation of the morphological rules of sted elimination and stress reduction in compound nouns WORD gt noun WORD gt noun particle WORD 1 s fuge is equal to 2 s fuge In this formulation compound nouns retain a single main stress only viz that of the initial component All other stresses are suppressed by the operator in the second WORD rule compare with the rules of double names in 3 1 above retaining the stress of the final component only All st d is eliminated from non final noun components by the operator The condition 1 s fuge is equal to 2 s fuge on the second WORD rule controls the fuge selection by feature unification In feature conditions symbol 1 refers to the first grammatical symbol after the arrow in casu noun while 2 refers to the 2nd symbol particle and so on O refers to the left hand symbol The present description gets the st d and stress patterns right even in multi compounds with more than two component nouns skovhest sg2Xw h2Esd sg2XwhEsd hesteskov h2Esd sg2Xw n2Esd0sgXw svinehundehus sv2i n h2un h2u s sv2i nOhunOhu s hussvinehund h2u s sv2i n h2un h2ussvi nOhun hundehussvin h2un h2u s sv2i n h2unOhussvi n Of course this does not conclude the description of the prosody of Danish compounds Many phenomena are still u2. The only permitted terminal symbol is WORD The enriched rule inventory allows for context free SYNTAX rules as in this classical example SYNTAX LINE gt NP VP NP gt N VP gt V NP N gt WORD V gt WORD In this grammar feature tests would be needed to impose categorical restrictions on the N and V rewritings 6 8 Stress reduction Any symbol terminal or not can be marked for stress reduction with the prefix minus sign Marking a terminal symbol means that the lexical stresses are removed from the phonetic form for that particular word If a non terminal symbol is marked all stresses will be removed from all words dominated by that symbol in the parse tree NP NAVNE NAVN NAVNE gt NAVN NAVNE gt NAVN NAVNE These SYNTAX rules removes all stresses from compound proper names except the final one cf 4 2 Stress reduction is also useful in MORPHOLOGY rules KOMPOSITUM gt subs FLERE FLERE gt subs FLERE FLERE gt subs These morphological rules notice the three dash arrows rewrite a compound noun as a string of stems all but the initial one being marked for stress reduction With a suitable lexicon file subs a word like plastiktandborstehar plastic tooth brush hair will be pronounced with the correct stress pattern 6 9 St d elimination Elimination of sted is similar to stress reduction using instead of Any combination of and is allowed Ex
3. orthographics to phonetics function calls the TtT interpreter translating your IL input to a phonetic string as specified in your lexicon and grammar rules This function is only available if all the following source files are found in your domain a syntax file a morphological file and one or more lexicon files Syntax rules morphology rules and lexicons entries must co refer in a carefully controlled fashion as specified below upload REQUIRES A VALID PASSWORD This function transfers the input in IA to your TtT domain There are three valid kinds of input A set of morphological rules A set of syntax rules A set of lexical entries The first line in IA must contain the filename Valid filenames are MORPHOLOGY SYNTAX 2279 lex 2277 are four lower case letters in a z i e not Examples of valid filenames also known as TtT identifiers MORPHOLOGY SYNTAX leds lex abcd lex voka lex Examples of invalid filenames frequently encountered MORFOLOGI MORFOLOGY morphology MORPHOLOGY ver 2 2 2009 SYNTAKS navne lex Verb lex cons voka LEX leds lex aben lex 6 3 Lexicon files The first line of a lexicon file is occupied by the lexicon identifier four letters plus lex Each of the following lines defines a lexical entry and must conform to this line format Ostring Pstring Featurestructure Ostring isa string of graphemes usually but not necessarily alphabetical letters a
4. Mozilla or Explorer The user must insert and upload three grammatical resources The TtT synthetic voice was developed by the project group Dansk Syntetisk Tale a consortium of two Danish universitites Copenhagen and Aalborg and a major Danish tele communication company TeleDanmark A S partly sponsored by the Danish Ministry of Research 1 a number of lexicon files specifying for each entry its e orthographic form e phonetic form e structural features if any 1 a file with morphological rules specifying how each input word segments into parts appearing in the lexicon files 2 a file with syntactic rule specifying how the input words relate to each other While we expand on the roles of the various grammatical resources in the following sections this section presents the TYT Workbench as seen from the web user s point of view M TtT workbench ver 1 Peter Juel Henrichsen CHOL CBS 7 File Edit Mew Go Bookmarks Tools Window Help LTTE TTT Tekst til Tale User id guests type guest Phone string ensdit2ud fC d2z taleNvisdig guests click phon2 way alive awake phon2way advanced Ge Eb Z ES d Figure 2 TtT Workbench basic version browser image Fig 2 shows the basic version The buttons alive and awake are for testing the connection between client and server This version allows the user to enter ready made strings of phones such as User input d2an sg aR s2an0li 2eg0 12
5. NOU N_UNIT indef NPrruuu gt NOUNj ger NPruy gt determiner NOUN UNIT 1 s definiteness is equalto 2 s definiteness NOUN UNIT inde g gt NOUN finder NOUN UNITfinder gt adjective NOU Nfindef O s definiteness is equalto 1 s definiteness In this section we present the TtT rules in a simplified form see section 6 for details Notice the condition on the second NOUN UNIT rule percolating the definiteness value from the adjective s to the NOUN UNIT via the feature condition 0 s definiteness is equal to 1 s definiteness while requiring the noun to be in indefinite form The grammar thus permits the full objects det nye hus the newa house and et nyt hus a newinaer house while excluding det nye huset well formed in Swedish and Norwegian but not in Danish We leave it as an exercise for the reader to compile the various rules appearing in this paper into a coherent TtT grammar upload it and listen to well shaped or bizarre sentences like Henning k ber hus Niels Henning rsted Pedersen k ber nyt hundehus Carl Carl Carl k ber den fine nye markmuseskibshushest 5 Concluding Remarks The TtT was created mainly as a tool for advanced students of speech technology and phonetics and as such it has served well for almost eight years now A follow up to this working paper is in preparation reporting on our use of this and other tools in the class room As experience has shown the TtT level of grammatic
6. Teleteknik 93 1 reprinted in Henrichsen 2001b 2007 Petersen N R amp Hansen P M 1994 Fundamental frequency resettings pauses and syntactic boundaries in read aloud Danish prose Acta Linguistica Hafniensia 383 400 Rischel J 1996 Frembringelse af syntetisk dansk tale Problemomr der Univ of Copenhagen reprinted in Henrichsen 2001b 2007 Thorsen N alias Gr nnum N 1978 An acoustical investigation of Danish intonation J of Phon 6 151 175 Gr nnum N alias Thorsen N 2005 Fonetik og Fonologi almen og dansk Copenhagen Akademisk Forlag Papers marked with concern the groundwork of the Danish synthetic voice developed in the project Dansk Syntetisk Tale
7. least one stressed syllable This is true even of words which are always subject to stress reduction when appearing in a phrase such as adverbials jo similar to y know and sgu a mild swearword or the infinitival particle at to Jo is thus lexicalized as 320 though it is always pronounced jo component has to be transmitted without error otherwise the whole name will fail to refer Each word thus has retained its main stress symbol 2 Niels Henning is considered as a single word since pronouncing Niels and Henning as two individual words n2els h2EneN would fail to signal the dash In contrast mode B is the one used when talking to someone who has at least heard of NH P before In this case the task is to simply establish a reference to an already known individual The three words now convey a single piece of information and this fact is prosodically encoded by the suppression of all lexical stresses except the final one The contrast between pronunciation modes A and B may be succinctly expressed in TtT terms Mode A simply amounts to mode A FULLNAME gt WORD name FULLNAME gt WORDrname FULLNAME In TtT grammars the symbol WORD functions as the link between the syntax rules where it plays the role of a terminal symbol and the morphological rules where it constitutes the top symbol In the rule set above mode A the first FULLNAME rule permits input consisting of a single name component
8. only e g Carl which will then be rendered exactly as specified in the lexical phone string including the sted as well as the main stress The mode B set differs from A at one point only viz the instruction to suppress the lexical stress and any sted from all non final name components mode B FULLNAME gt WORD name FULLNAME gt WORDrname FULLNAME Choosing between mode A and mode B on the fly may require a good deal of conversational practise to master but the student who knows the formal rules of pronunciation at least does not stumble around in the dark 4 3 The Case of Compound Nouns Consider a set of monosyllabic Danish common nouns hest hund svin mus hus skov mark skib horse dog pig mouse house woods field ship They all share the same basic phonetic structure a pre vocalic consonant group followed by a full vowel and a post vocalic consonant group In English most nouns of this structure form compounds by simple concatenation with little or no phonetic alteration of the components In Danish in contrast compound noun formation has quite complex effects on the phonetic rendering In the Danish compound noun as a general rule all main stresses are suppressed except for the first In some cases this is all that happens mark m2A g mus n2u s markmus m24 gmu s However usually further prosodic transformations are in play The st d of the first noun component may be elimi
9. the lexicon file with that identifier The top symbol of the MORPHOLOGY grammar is WORD by definition This symbol must be re written in other words it must occur as the left hand symbol of at least one rule 6 6 Jump The nill symbol may replace a terminal symbol in any rule resulting in a so called jump rule corresponding to a jump arc in a finite state network Jumps do not involve lexical lookup they can sometimes be used for smoother grammar design Here is an example of a grammatically well formed MORPHOLOGY file MORPHOLOGY WORD gt prae MORE WORD gt prae STILLMORE MORE gt stop STILLMORE gt prae STILLMORE STILLMORE gt stop STILLMORE gt MORE Notice that jumps must be used cautiously Of course recursive rules including will cause endless looping but you should also be aware of mutual recursiveness among two or more rules as in A gt B B gt A The use of feature structures and feature tests will be introduced later 6 7 The SYNTAX file The SYNTAX rule format is similar to the MORPHOLOGY format but with a few important differences The rewriting arrow is gt dash dash greater than The permitted rule inventory is larger NONTERM gt TERM NONTERM gt TERM NONTERM NONTERM NONTERM TERM NONTERM gt NONTERM NONTERM again each rule may be followed by a feature test The top symbol is LINE
10. Ed dansk er sandelig ikke let Danish is certainly not easy Words may be comma separated The commas have no influence on the acoustic rendering but they tend to make phonetic strings more readable 7 The TtT Workbench phonetic inventory is based on the Danish SAMPA www phon ucl ac uk home sampa a many to one mapping of the IPA International Phonetic Alphabet on the Danish sound inventory Since certain SAMPA symbols are inconvenient for use with regular expressions as in the TtT server scripts and for transfer over the Internet used by the TtT web interface we use an alphanumeric SAMPA mapping The complete TtT phone inventory can be consulted in section 7 2 is tryk main stress is sted a quick glottal contraction is vokalforl ngelse vowel prolongation is the full vowel in e g vil will shall want to is the full vowel in e g vel probably is the full vowel in e g male paint is schwa as in e g male paint On WO TtT phonetics does not include secondary stress See the full TtT phone table in sect 7 Table 1 TtT phone table summary By pressing the button phon2way the client transmits his input to the server which in turn returns the sound file in wav format produced from the phone string Most browsers will then allow the user to just click on the link on the answer page in order to listen to the sound file For pedagogical reasons the TtT server application
11. The CBS Text to Speech Workbench Peter Juel Henrichsen Copenhagen Business School pjh isv cbs dk Abstract This working paper presents the CBS text to speech tool colloquially known as the TtT Tekst til Tale The tool is intended for training of university level students especially linguists training for a degree in speech technology and visiting foreign students wanting to improve their spoken Danish The TtT is operated through a simple www based user interface Using the TtT requires basic skills in formal grammar writing but no knowledge on other aspects of artificial voice development such as phonetic acoustic quantification prosodic modelling and signal generation The paper includes a user manual 1 Introduction The TtT workbench Text til Tale text to speech is a simple to operate web based tool for creating Danish synthetic voices The TtT is aimed mainly at three kinds of university level students the computational linguist training for a degree in speech technology the visiting foreign student learning Danish and the phonetician The tool is available for non commercial teaching and research purposes The tool is www based Apart from the obvious practical advantages this has allowed us to obtain permission to use proprietary speech technological software thanks to the low risk of accidental exposure User 1 The speech technology student University level courses in speech technology often face a pedagogical prob
12. adopts a rather conservative style of feed back rejecting with a comment any irregular phone string Examples of phone strings rejected e Strings beginning with a semivowel e g R J or w in conflict with the Danish phonotax e Strings with zero instances of symbol 2 main stress any utterance must contain a stressed syllable in order to be pronounceable e Strings with illegal st d Only two sted loci are permitted viz immediately after a long vowel as in ben b2e n and immediately after a short vowel voiced consonant as in bind b2en It is recommended to have students work with version 1 until they master the TtT phone inventory before introducing them to version 2 M TET workbench ver 2 Peter Juel Henrichsen CHOL CBS File Edit Mew Go Bookmarks Tools Window Help Tekst til Tale ORPHOLOGY ORD gt name wider narrower DEFAULT higher lower upload dir ort2phon phon2wav Figure 3 TtT Workbench advanced version browser image Each button in version 1 also appears in version 2 advanced accompanied by a few more upload for uploading of resource files dir for access to uploaded files and ort2phon for generating phone strings from orthographic input When the student has completed his language model by uploading valid lexicon files together with a complete and consistent grammar the ort2phon button becomes active Pushing the button will now convert an orthographic inp
13. al abstraction makes it suitable for students who have already passed basic courses on phonetics and formal grammar and who have some practical experience in rule writing in the Chomskian tradition Concerning TtT as a tool for L2 learning we have less experience but high hopes We have touched on a number of rules of Danish pronunciation rules that may be specified either in fully formal terms as in the toy grammars in section 3 or in semi formal paraphrase as exemplified in Figure 4 below 1 Pronunciation of non compound names is as specified in the lexicon 2 In compound names N1 N2 double names the lexical stress is always eliminated from NI 3 In compound names N N2 double names any st d is eliminated from NI 4 Pronunciation of full personal names NI N2 Nm has two modes call them A and B e In mode A name unknown to the hearer all names are pronounced as specified by their morphology cf 1 3 above e In mode B name known to the hearer all lexical stresses and st d are eliminated except in the final name component which is pronounced as specified by its morphology Figure 4 Rules of Danish prosody The rules governing the unit stress are actually more complex For instance the verb phrase k ber mange huse buys many houses does exhibit the unit stress while k ber nogle huse buys some houses does not this difference is not predicted by our example grammar both VPs
14. amples of MORPHOLOGY and SYNTAX rules COMPOUND gt noun noun as in husmus PP P NP as in p den ene side VGROUP VAUX PTC as in o skulle have v ret spist 6 10 Feature tests Any rewriting rule in SYNTAX and MORPHOLOGY may be followed by an optional feature test LHSO arrow RHS1 Feature test LHSO arrow RHS1 RHS2 Feature test where LHSO is the left hand side symbol to be rewritten RASn the right hand symbols and arrow either gt or gt Notice that the left hand side symbol is indexed with 0 and the right hand symbols with I or 2 These indices are used for reference in the feature test The feature test consists of any number of conditions written one at a line This rewriting rule thus has 4 conditions VP gt VAUX VINF l type lt aux form finit 2 valens intrans 2 form infinit The infix operator lt means has the value and ix fea refers to feature fea of symbol ix Only lower case letters in a z are allowed in features and values The first condition 1 type aux thus requires that the feature type ofthe right hand symbol VAUX has the value aux Incase type is uninstantiated has no value set by the time the rule is called type will be set to aux as a side effect of the evaluation this value cannot be changed again If type was already instantiated the condition is fulfilled only if its value is equal to aux The other infix operator used in fe
15. ature tests is lt gt less than greater than It means has the same value as Its use is best illustrated by an example NP gt DET CN l gender 2 gender O gender lt gt 1 gender l num lt gt 2 num 0 num lt gt l num This NP rule is conditioned on the congruence of the determiner and the noun wrt gender and number Et hus anzu sna housevev snc og en cykel aur sva bikeur sno are wellformed noun phrases hence accepted by the grammar while et cykel angu sua bikeur snc and en cykler aur sne bikesur pLur are rejected If the CN is underspecified for gender as is the case for certain Danish nouns like fond foundation the gender feature of the NP will be properly instantiated as a side effect of the lt gt application in the condition 0 gender lt gt 1 gender A feature test is fulfilled if and only if each of its conditions is met If the test fails the rewriting rule does not apply VP gt V OBJECT 2 unitstress lt yes VP gt V OBJECT 2 unitstress lt no OBJECT gt DET N O unitstress lt no OBJECT gt N l defness lt yes O unitstress lt no OBJECT gt N l defness lt indef O unitstress lt yes This example shows how the 0 index can be used for coding a phrase according to a syntactic property As discussed in sect 4 4 the application of the Danish unit stress in verb phrases is partly a function of the topology of the grammatical objec
16. consistently between the two e g in words of foreign origin meter motor cancer censor Alexander Hektor z is almost always prolonged notable exceptions being words using the letter as a morpheme A4 ark ABC bog A38 a and C are rarely prolonged except in interjections like ja s 0 schwa is never prolonged or stressed neither does it licence st d Table 2 Vowel symbols in TtT with nearest IPA equivalents PRAAT encoding Vowel Phonetic variants As in TtT IPA Short Long St d Stress a p md fe fe ES EEN dete det fe fees nase hes diy Tz ree fame fafa e femte a RE ee ba ben a e porse re per fontonet Dx roe ET a fo f f Jugend E E po ne mt EE fi be i e e e bomo ae for he vi Fo fue fog infinite Cecima Table 3 Consonant symbols in TtT with nearest IPA equivalents PRAAT encoding Consonant Syllabic positions and sted licencing Other Post As in TtT IPA Inital Stod base prevoc vocalic de at eni EN ds p dhe EERE TS SE SS SC E Oo ea EN A I 7 ES eee ces cs uH Hr Hr AI uH i Apo T T P T o 7 slu l l dt ERE EEN 72 K 5 H Q Sa QIo T T i 7 1 Further conventions A full vowel is any vowel with the exception of
17. ed L2 learner has internalized already This fact is however not always exploited by language teachers Academic level courses are typically based on traditional grammar books which in general do not consider prosodic fine structure to belong to the realm of real grammar In so far as pronunciation is studied at all beyond the word level this is often done unsystematically on a trial and error basis with the highly idealized phonetic annotations of a standard dictionary as the only technical reference It is then left to the learner s musical ear to discover the regularities of the grammatically governed stress reductions sted eliminations vowel shortenings etc The sted has been presented in the literature as a quick glottal stop but is actually better described as an instance of creaky voice Like the tonemes of Swedish and Norwegian of which the sted is actually a reminiscent it has the status of a meaning separating feature in the structuralist sense B nder peasants and b nner beans are thus distinguished phonetically by the sted only User 3 The phonetics student The phonetics student is actually preoccupied with the very same phenomena as the L2 learner however at another level of abstraction Whereas the latter needs training in the practical application of the Danish rules of pronunciation the former must learn how to design the very same rules The phonetics student thus begins his TtT experience with an empty r
18. he ubiquitous vocoid cluster reductions det er en de n der er jo dAo all conspire to frustrate the L2 learner The many foreign CBS students who are required to learn Danish within one or a few semesters thus face a problem Der til medh s w rdas de icke Dansk er et Danish is not a heller att talla som annat folck nordisk sprog language uthan tryckia ordhen fram lika som med plattyske it is a throat the willia hosta och synas endeles udtalevaner disease medh flitt forwendhe ordhen i strupan for n de komma framn J rgen Rischel anonymous Hemming Gadh 1510 Two specific aspects of Danish pronunciation are perceived by L2 learners as particularly hard to master the stress assignment and the stod Standard Danish Typical L2 speaker errors hold k ft h old k ft shut up skal vi drikke k affe sk al vi drikke k affe should we drink coffee han vil ikke g hj em han vil ikke g hjem he won t go home h un den i h undehu set h unden i h undh uset the dog in the dog house is tryk main stress is sted a quick glottal contraction Such errors are actually easy to pinpoint and eliminate The distribution of stress sted vowel shortening and related phonetic features are largely regular and can hence be expressed in ordinary grammar rules Such rules are quite easy to state and test in TtT Many of the regularities of Danish prosody can be described with reference to grammar rules that the advanc
19. iniscent of the formalism defined by Peter Molb k Hansen for project Danish Synthetic Speech The present TYT Workbench is however a complete rewrite and does not contain any code from the DSS project The groundwork underlying DSS see References titles marked with To most non native speakers of Danish the enhedstryk is one of the prosodic puzzles that remain unsolved However the most important rules governing the unit stress are not difficult to express or internalize Compare the pronunciation of the transitive verb k ber buys in the following sentences only the verb is marked for stress and st d Niels k o ber et hus Niels buys d house Niels k ber et nyt hus Niels buys a new house Niels k e ber huset Niels buys the house Niels k o ber husene Niels buys the houses Niels kober hus Niels buys a house Niels k ber nyt hus Niels buys a new house Niels k ber huse Niels buys houses is main stress is st d Two conditions on the grammatical object must be fulfilled in order for the object to trigger a stress reduction enhedstryk on the main verb e indefinite form e no overt determiner Such enhedstryk triggering objects are called bare while non triggering objects are called full Hence enhedstryk is described in two rules VP gt verb N Duare VP gt verb NPiun As seen the object NPs must be marked for object type bare or full The marking is done in the NP rules NP bare gt
20. lem Even for advanced linguistics students who have already passed courses on phonetics formal grammar and natural language processing reaching the level of professional working knowledge within a single semester can be extremely demanding Speech technological development be it ASR automatic speech recognition or TTS text to speech synthesis requires expertise in an unusually broad range of disciplines from psycho acoustics and advanced statistics to formal grammar and lexical phonology Good learning results depend heavily on a carefully balanced syllabus backed up by designated tools A number of such tools have been developed at CBS based on seven years of teaching experience The TtT is the prime example By hiding away most of the technical complexity of a full blown TTS system behind a simple user interface the tool allows the student to concentrate on linguistically relevant aspects of synthetic voice development The student can thus design rewriting grammars for those aspects of morphology compositional phonetics and prosody that are relevant in a production system for rule based artificial speech User 2 The foreign student learning Danish Danish spoken language is notoriously difficult to learn Not only is the relationship between the spelling and the sound of words often opaque the vast vowel inventory greatest among all European tongues the subtle rules of sted and stress application compound formation and t
21. m asleep the TtT application is currently off line I m lost you didn t provide a user id or a password I m confused your password is formally ill formed I m suspicious your password is formally well formed but incorrect I m stoned the cpu is currently over exposed or facing other problems I m flabbergasted the query is nonsensical did you fiddle with the URL phon2wav Given a well formed phonetic string in IL this function will return a sound file in wav format generated as specified in your input The phonetic string must be In HTML terms the input line and input area correspond to form elements lt input type text gt and lt textarea gt respectively preceded and followed by the separator symbol comma Optionally may be inserted at any point in the phonetic string e g between words for improved legibility This has no effect on the produced sound file Examples of valid input en t2iq C aR m2e C En en 2Emq C en tigre er mere end en femore 2eNOn k2EnC n2ad0n ingen kender natten nels hEneN QRsdED p2e dCsOn Niels Henning Orsted Pedersen dir REQUIRES A VALID PASSWORD This function returns a list over all files currently held in your TtT domain The filenames are clickable On click the contents of a file are returned into the IA for you to modify and re submit cf Upload below ort2phon REQUIRES A VALID PASSWORD The ort2phon
22. names separated by dash The operators minus and question mark in the second WORD rule marks the first name component for stress reduction and st d elimination respectively The terminal symbol dash in the 2nd WORD rule refers to a new lexicon file with just a single entry LEXFILE dash in which the symbol means no phonetic output Given an input like Niels Henning this TtT grammar thus produces the phone string nelsh2EneN Carl Henning and Niels Carl are rendered as kA 1h2EneN and nelsk2A 1 respectively Notice that the sted is eliminated in the former case only which is correct 4 2 The Case of Niels Henning rsted Pedersen You can tell from the pronunciation of a Danish name whether the speaker believes the hearer to be acquainted with the person referred to Compare these two renderings of the name Niels Henning rsted Pedersen cf phone table in sect 7 A nelsh2EneN 2xRsdED p2e DCs0n B nelshEneN xRsdED p2e DCs0n Mode A is the one NH P would probably use in a situation where he was submitting his name to say a social security agent NH P might even insert a small pause at the points of the commas to make the segmentation of his name easier to grasp The what is your name dialogue is clearly a low redundancy situation where each name Any lexicalized phonetic string must contain at least one stress symbol 2 since any Danish word pronounced in isolation has at
23. nated as in skov sg2Xw svin sv2i n skovsvin sg2Xwsvi n X is the full vowel in kork In addition a long vowel may be shortened hus h2u s mus m2u s husmus h2usmu s To this comes the fuge glue particle a phonetic element often required to glue together the components in compound nouns and other categories as well In modern Danish only two kinds of fuge remain productive the schwa fuge and the s fuge hund h2un hus h2u s hundehus h2unOhu s skib sg2i b hund h2un skibshund sg2ibshun Il It is not possible in general to determine the fuge selection stod elimination and other phonetic transformations of a noun from its lexical phone string alone In Danish compounding is a complex interaction of lexical morphological and prosodic rules Advanced Danish dictionaries usually do contain some information on compounding but it is almost impossible even for the advanced student to verify his pronunciation of arbitrary compound nouns on the basis of lexical look ups alone Consider a toy example a TtT grammar for compound noun formation As seen lexicon entries now include information on fuge selection LEXFILE noun hest h2Esd fuge schwa hund h2un fuge schwa svin sv2i n fuge schwa mus m2u s fuge schwa hus h2u s fuge nill skov sg2Xw fuge nill mark m2A g fuge nill skib sg2i b fuge s LEXFILE particle e 0 fuge schwa S S fuge s
24. nd numerals defining an orthographic form usually but not necessarily a morpheme or a full word Space characters are not allowed in Ost ring while underscores are Pstring isa well formed phonetic form consisting exclusively of the phones defined in the TtT phone table sect 7 in a formally valid sequence Notice that some of the symbols required in input to Ort2Phon are not used in lexicon entries Featurestructure isa sequence of grammatical feature value pairs optional Each pair has the form feature value Both symbols are written with lower case letters in a z Examples gnd neutrum num plur definiteness indef This particular feature value set could come from a lexical entry like bern children coded for neuter gender and plural indefinite form The complete lexical entry would then be b rn b2xR n gnd neutrum num plur defin ubest Any number of space characters may be inserted between the lexical segments ostring pstring and feature value pairs but nowhere else Blank lines are permitted everywhere except as the first line in the file which must carry the filename Example of a full lexicon file with the TtT identifier abcd abcd lex S S y 2y 12 tcl sk sg du d2u d d2q form inf diatese active Mortensen m20 dO0nsOn cat pn case neu A38 2y mC gnd ut num sing taste sour 6 4 The nill symbol The hash character 1 serves as a general nill symbol It may replace an ostring to the effect that
25. nexplained e g why mus and hus have their vowels shortened when compounding while svin and mark do not compare musesvin m2usOsvi n and svinemus sv2i n0mu s Neither have we touched on the fact that some nouns selecting an s fuge retain their st d in compounds contrary to the predictions of the toy grammar above such as stat in statsskov sd2z dssgXw and avl in avlshest 2Aw 1shEsd resp state forest and breeding horse Such problems can in fact be tackled in the TtT formalism however this is not the place to elaborate 4 4 The Case of the Unit Stress In the Danish main clause certain types of verbal complements trigger a stress reduction on the main verb This phenomenon is known as enhedstryk unit stress From a semantic point of view its effect is akin to that of the B rule for personal names cf 3 2 Intuitively speaking applying a unit stress creates a tighter more unit like semantic constituent to be analyzed en bloc rather than en detail Often VPs with unit stress may be substituted by a single intransitive verb in contrast to VPs retaining the stress on the main verb The verb phrase at kobe varer to buy goods exhibits the enhedstryk and it can often be replaced by at handle to shop in contrast at kobe en vare to buy an item or at kobe varerne to buy the goods have no enhedstryk and no similar one verb equivalent The numbering convention in feature conditions is rem
26. ning information on the query processing normal abnormal comments and or error messages and most importantly the outcome of the grammatical analysis including a variety of morphological and syntactic details Many parts of the parsing information are clickable such as the individual morphological analyses On click these results will be transferred to the TtT main page and inserted in the IL for easy testing with the ort2phon and phon2wav buttons In phonetic strings areas which have been marked for stress reduction as an effect of the parsing are marked with red color while areas with st d elimination are shown in italic font 7 TtT phone definitions Phonetic symbols permitted in TtT input and lexical entries The TtT phone inventory is inspired by the Danish SAMPA alphabet with small modifications cf http www phon ucl ac uk home sampa TtT phone symbols are all alphabetical while diacritics sted vowel prolongation etc are not Digits are reserved for current and future degrees of stress e g 2 for lexical main stress The TtT vowel inventory includes only one symbol for the three vowels in e g Sommerslot s2CmCs1Cd The phonological distinction between the full vowel in nok n2Cg and the vowel found in morphological endings sometimes called schwa 2 as in baner b2z nC l ngere 12ENCC is thus collapsed in TtT phonetics In our experience students and experts alike find it very hard to distinguish mm
27. no orthographic input is read ora pst ring no phonetic symbols produced or even both representing the jump arc known from finite state grammar theory Warning Use the nill symbol very cautiously especially when engaged as ostring Careless use may lead to endless loops 6 5 The MORPHOLOC Y file The file named MORPHOLOGY contains the morphological rules i e those describing the relations between the parts of a token usually a word The first line of the file must be the identifier MORPHOLOGY followed by lines with rules in one of these formats NONTERM gt TERM Feature test NONTERM TERM NONTERM Feature test where NONTERM is a non terminal symbol TERM isa terminal symbol Feature test is used for placing restrictions on the applicability of the rule Feature tests are optional The rewriting arrow gt is written with three dashes and a greater than sign it must be preceded and followed by one or more space characters Readers trained in formal grammar will recognize that MORPHOLOGY rules are finite state as the rewriting arrow must be followed by a terminal symbol as opposed to the context free rules of the SYNTAX file Further specifications Non terminal symbols are written with upper case letters in A Z i e not using Examples POSTVOK FULDVOKAL KLUSIL X Terminal symbols are written with four lower case letters in a z not Each terminal symbol refers to
28. obviously containing overt determiners Memorizing and utilizing prosodic rules at this level of abstraction is extremely demanding on the language student indeed this pedagogical practice has long been deprecated On the other hand abandoning rules altogether and relying solely on the student s intuition is not ideal either denying him the insight that prosody is to a large extent rule governed An attractive compromise could be to allow the student to play around with the formal principles of pronunciation while listening to the prosodic effects of his grammatical choices Working with formal grammar rules develops a high level of grammatical consciousness As a supplement to traditional conversational training CALL tools based on speech technology may therefore be of assistance to the advanced student Even if quite a few ambitious learners have trained with TtT over the years the current version can hardly be marketed as an off the shelf pedagogical tool for L2 teaching To serve that purpose the user interface would have to be developed graphically and accompanied by structured exercises low tech manuals and even didactic principles It is our hope that some reader of this paper would take up that challenge 6 User s manual 6 1 Input The TtT Workbench tekst til tale text to speech is available in two versions basic and advanced The latter comprises all the features of the former plus several more while the opposite i
29. s not true Therefore we only present the advanced version here There are two ways of uploading information to the advanced version of the TtT Workbench hereafter TtT for short an input line IL and an input area IA The user can insert one line of text into the IL and any number of lines into the IA While there are no formal restrictions to the amount of input TtT is not a convenient tool for building large scale applications IL data of up to 100 bytes and IA up to a few thousand are thus reasonable practical limits 6 2 The buttons and their functions Most TtT functions return a query status in the form of a string of numerals When the string consists of Os only this is a sign that the query has been 1 evaluated and found to be well formed and ii processed without technical problems Any non 0 digits refer to error types to be reported if the technical staff at CBS are contacted Pushing this button should get you the reply I m alive as a signal that the TtT web server is in service and your own web connection is effective Please operate this button before reporting any error to the TtT administrators This function does not require a password awake This button is for validating your TtT id and TtT password The reply should be I m ready together with an acknowledgement of your system id otherwise your id is not accepted or your TtT account has been blocked Other possible replies include I
30. schwa symbol 0 As mentioned the TtT conventions thus does not support the schwa 2 analysis of the vowel C in words like venner v2EnC l ngere 12ENCC proposed by some phonologists TtT uses the colon for vowel prolongation penne p2En0 vs p ne p2E n0 Any full vowel may be prolonged see table 2 Main stress is marked with 2 prefixed to a full vowel e g 2i Any full vowel can carry a main stress St d is annotated with the exclamation mark It may occur in one of two different positions 1 just after a prolonged vowel and ii just after one of these consonants mn 1 JRwDN if preceded by a short full vowel see table 3 Examples vej v2AJ but not v2A J hegn h2AJ n but not h2A Jn h2AJn p n p2E Ln but not p2E n pen p2En but not p2E n References Henrichsen P J 2001a Transformation based learning of Danish stress assignment Proceedings of Eurospeech 01 Henrichsen P J 2001b Kompendium i Taleteknologi med s rlig v gt p syntese af dansk talesprog Copenhagen Business School Press Henrichsen P J 2007 Tekster til Taleteknologi med s rlig v gt p syntese af dansk talesprog Copenhagen Business School Press Holtse P et al 1991 IAAS TFL Speech Synthesis Project report 3 Cph Work Pap in Ling 1 146 168 reprinted in Henrichsen 2001b 2007 Holtse P et al 1993 Syntetisk tale en introduktion
31. t rather than properties of the participating lexemes The decision concerning stress reduction in the VP rule thus depends on information which none of the lexemes can provide The suggested solution uses a feature unitstress defined exclusively in the SYNTAX file to mediate between the VP and OBJECT constituents The underscore symbol is used for underspecified values in the lexicon This can often be useful for lexemes with underspecified feature values trist tr2isd gnd ut num sng def no adj sad trist tr2isd gnd neu num sng def no bl t b12Cd gnd neu num sng def no adj blue bl bl2c gnd ut num sng def no bl bl2c gnd neu num plur def no bl bl2c gnd ut num plur def no bl bl2c gnd neu num sng def yes bl bl2c gnd ut num sng def yes bl bl2c gnd neu num plur def yes bl bl2c gnd ut num plur def yes Using the underscore these rather redundant entries can be replaced by trist tr2isd gnd num sng def no blat b12Cd gnd neu num sng def no bl bl2c gnd neu num plur def no bl bl2c gnd ut num def no bla bi2e gnd num def yes 6 11 File deletion Use the IL for deleting unwanted files from your TtT directory If you wish to abolish your SYNTAX file simply type SYNTAX followed by nothing and push Upload Similarly for MORPHOLOGY and lexicon files 6 12 Output Output from TtT is presented in the form of an html page contai
32. ule directory to be gradually filled up over the semester whereas the L2 learner meets TtT in a fully instantiated version ready for use 2 Presenting the TtT At its core the TtT Workbench has a Danish synthetic voice the voice known as Carsten owned by Mikro Veerkstedet A S A web interface provides the user of the workbench with restricted access to the speech synthesizer allowing him to define the lexical descriptions and grammar rules governing the artificial voice while hiding away all further technical details Thus the user has full control over the linguistic parameters but at the same time he has no influence on the technical aspects of the processing such as the parsing strategy the calibration of the prosodic model the calculation of the frequency envelope of the utterance the timing of the phone sequence or the encoding of the sound signal The user may thus concentrate on writing lexicon entries and grammar rules and forget about everything else Figure 1 shows the logical parts of the speech engine To the user of the TtT Workbench only the two uncoloured databases at the bottom are transparent while all the other modules function as a black box TEXT INPUT SOUHD OUTPUT Text Prosody Signal E gt 22 ET Le Grammar Rules Figure 1 TtT Workbench logical components 2 1 The Web Interface The web based interface allows the user to control the speech engine using a standard browser e g
33. ut string to a phonetic output string which in turn may be converted into a sound file using the phon2wav button as already explained 4 Pronunciation patterns expressed in TtT rules In the following sections we present a number of grammatical case studies suggesting how problem areas of Danish pronunciation could be charted using the TtT Workbench The cases are organized so as to progressively introduce the basic features of the TtT Workbench Henrichsen 2001b 2007 4 1 The Case of Given Names and Double Names Even if double names are perhaps not as frequent in Denmark as in Sweden or Holland given names like Anne Sofie and Carl Erik are far from unusual It is a general rule of Danish pronunciation that such double names should be rendered with one stress only always located in the second of the connectednames Also all instances of sted in the first name are to be eliminated Such rules are easily reformulated in TtT terms Consider first a TtT lexicon file of personal names The file id is name LEXFILE name Niels n2els Henning h2EneN Carl k2A 1 Each lexical entry consists of an orthographic form and its associated phonetic rendering The morphological rules in a separate file make reference to the lexicon file or files in case there is more than one WORD gt name WORD gt name dash name The first rule permits words consisting of one personal name only stress retained while the second rule analyzes double
Download Pdf Manuals
Related Search