This is festival.info, produced by Makeinfo version 3.12h from festival.texi. This file documents the `Festival' Speech Synthesis System a general text to speech system for making your computer talk and developing new synthesis techniques. Copyright (C) 1996-2001 University of Edinburgh Permission is granted to make and distribute verbatim copies of this manual provided the copyright notice and this permission notice are preserved on all copies. Permission is granted to copy and distribute modified versions of this manual under the conditions for verbatim copying, provided that the entire resulting derived work is distributed under the terms of a permission notice identical to this one. Permission is granted to copy and distribute translations of this manual into another language, under the above conditions for modified versions, except that this permission notice may be stated in a translation approved by the authors. File: festival.info, Node: Emacs interface, Next: Phonesets, Prev: XML/SGML mark-up, Up: Top Emacs interface *************** One easy method of using Festival is via an Emacs interface that allows selection of text regions to be sent to Festival for rendering as speech. `festival.el' offers a new minor mode which offers an extra menu (in emacs-19 and 20) with options for saying a selected region, or a whole buffer, as well as various general control functions. To use this you must install `festival.el' in a directory where Emacs can find it, then add to your `.emacs' in your home directory the following lines. (autoload 'say-minor-mode "festival" "Menu for using Festival." t) (say-minor-mode t) Successive calls to `say-minor-mode' will toggle the minor mode, switching the `say' menu on and off. Note that the optional voice selection offered by the language sub-menu is not sensitive to actual voices supported by the your Festival installation. Hand customization is require in the `festival.el' file. Thus some voices may appear in your menu that your Festival doesn't support and some voices may be supported by your Festival that do not appear in the menu. When the Emacs Lisp function `festival-say-buffer' or the menu equivalent is used the Emacs major mode is passed to Festival as the text mode. File: festival.info, Node: Phonesets, Next: Lexicons, Prev: Emacs interface, Up: Top Phonesets ********* The notion of phonesets is important to a number of different subsystems within Festival. Festival supports multiple phonesets simultaneously and allows mapping between sets when necessary. The lexicons, letter to sound rules, waveform synthesizers, etc. all require the definition of a phoneset before they will operate. A phoneset is a set of symbols which may be further defined in terms of features, such as vowel/consonant, place of articulation for consonants, type of vowel etc. The set of features and their values must be defined with the phoneset. The definition is used to ensure compatibility between sub-systems as well as allowing groups of phones in various prediction systems (e.g. duration) A phoneset definition has the form (defPhoneSet NAME FEATUREDEFS PHONEDEFS ) The NAME is any unique symbol used e.g. `mrpa', `darpa', etc. FEATUREDEFS is a list of definitions each consisting of a feature name and its possible values. For example ( (vc + -) ;; vowel consonant (vlength short long diphthong schwa 0) ;; vowel length ... ) The third section is a list of phone definitions themselves. Each phone definition consists of a phone name and the values for each feature in the order the features were defined in the above section. A typical example of a phoneset definition can be found in `lib/mrpa_phones.scm'. Note the phoneset should also include a definition for any silence phones. In addition to the definition of the set the silence phone(s) themselves must also be identified to the system. This is done through the command `PhoneSet.silences'. In the mrpa set this is done by the command (PhoneSet.silences '(#)) There may be more than one silence phone (e.g. breath, start silence etc.) in any phoneset definition. However the first phone in this set is treated special and should be canonical silence. Among other things, it is this phone that is inserted by the pause prediction module. In addition to declaring phonesets, alternate sets may be selected by the command `PhoneSet.select'. Phones in different sets may be automatically mapped between using their features. This mapping is not yet as general as it could be, but is useful when mapping between various phonesets of the same language. When a phone needs to be mapped from one set to another the phone with matching features is selected. This allows, at least to some extent, lexicons, waveform synthesizers, duration modules etc. to use different phonesets (though in general this is not advised). A list of currently defined phonesets is returned by the function (PhoneSet.list) Note phonesets are often not defined until a voice is actually loaded so this list is not the list of of sets that are distributed but the list of sets that are used by currently loaded voices. The name, phones, features and silences of the current phoneset may be accessedwith the function (PhoneSet.description nil) If the argument to this function is a list, only those parts of the phoneset description named are returned. For example (PhoneSet.description '(silences)) (PhoneSet.description '(silences phones)) File: festival.info, Node: Lexicons, Next: Utterances, Prev: Phonesets, Up: Top Lexicons ******** A _Lexicon_ in Festival is a subsystem that provides pronunciations for words. It can consist of three distinct parts: an addenda, typically short consisting of hand added words; a compiled lexicon, typically large (10,000s of words) which sits on disk somewhere; and a method for dealing with words not in either list. * Menu: * Lexical entries:: Format of lexical entries * Defining lexicons:: Building new lexicons * Lookup process:: Order of significance * Letter to sound rules:: Dealing with unknown words * Building letter to sound rules:: Building rules from data * Lexicon requirements:: What should be in the lexicon * Available lexicons:: Current available lexicons * Post-lexical rules:: Modification of words in context File: festival.info, Node: Lexical entries, Next: Defining lexicons, Up: Lexicons Lexical entries =============== Lexical entries consist of three basic parts, a head word, a part of speech and a pronunciation. The headword is what you might normally think of as a word e.g. `walk', `chairs' etc. but it might be any token. The part-of-speech field currently consist of a simple atom (or nil if none is specified). Of course there are many part of speech tag sets and whatever you mark in your lexicon must be compatible with the subsystems that use that information. You can optionally set a part of speech tag mapping for each lexicon. The value should be a reverse assoc-list of the following form (lex.set.pos.map '((( punc fpunc) punc) (( nn nnp nns nnps ) n))) All part of speech tags not appearing in the left hand side of a pos map are left unchanged. The third field contains the actual pronunciation of the word. This is an arbitrary Lisp S-expression. In many of the lexicons distributed with Festival this entry has internal format, identifying syllable structure, stress markigns and of course the phones themselves. In some of our other lexicons we simply list the phones with stress marking on each vowel. Some typical example entries are ( "walkers" n ((( w oo ) 1) (( k @ z ) 0)) ) ( "present" v ((( p r e ) 0) (( z @ n t ) 1)) ) ( "monument" n ((( m o ) 1) (( n y u ) 0) (( m @ n t ) 0)) ) Note you may have two entries with the same headword, but different part of speech fields allow differentiation. For example ( "lives" n ((( l ai v z ) 1)) ) ( "lives" v ((( l i v z ) 1)) ) *Note Lookup process:: for a description of how multiple entries with the same headword are used during lookup. By current conventions, single syllable function words should have no stress marking, while single syllable content words should be stressed. _NOTE:_ the POS field may change in future to contain more complex formats. The same lexicon mechanism (but different lexicon) is used for holding part of speech tag distributions for the POS prediction module. File: festival.info, Node: Defining lexicons, Next: Lookup process, Prev: Lexical entries, Up: Lexicons Defining lexicons ================= As stated above, lexicons consist of three basic parts (compiled form, addenda and unknown word method) plus some other declarations. Each lexicon in the system has a name which allows different lexicons to be selected from efficiently when switching between voices during synthesis. The basic steps involved in a lexicon definition are as follows. First a new lexicon must be created with a new name (lex.create "cstrlex") A phone set must be declared for the lexicon, to allow both checks on the entries themselves and to allow phone mapping between different phone sets used in the system (lex.set.phoneset "mrpa") The phone set must be already declared in the system. A compiled lexicon, the construction of which is described below, may be optionally specified (lex.set.compile.file "/projects/festival/lib/dicts/cstrlex.out") The method for dealing with unknown words, *Note Letter to sound rules::, may be set (lex.set.lts.method 'lts_rules) (lex.set.lts.ruleset 'nrl) In this case we are specifying the use of a set of letter to sound rules originally developed by the U.S. Naval Research Laboratories. The default method is to give an error if a word is not found in the addenda or compiled lexicon. (This and other options are discussed more fully below.) Finally addenda items may be added for words that are known to be common, but not in the lexicon and cannot reasonably be analysed by the letter to sound rules. (lex.add.entry '( "awb" n ((( ei ) 1) ((d uh) 1) ((b @ l) 0) ((y uu) 0) ((b ii) 1)))) (lex.add.entry '( "cstr" n ((( s ii ) 1) (( e s ) 1) (( t ii ) 1) (( aa ) 1)) )) (lex.add.entry '( "Edinburgh" n ((( e m ) 1) (( b r @ ) 0))) )) Using `lex.add.entry' again for the same word and part of speech will redefine the current pronunciation. Note these add entries to the _current_ lexicon so its a good idea to explicitly select the lexicon before you add addenda entries, particularly if you are doing this in your own `.festivalrc' file. For large lists, compiled lexicons are best. The function `lex.compile' takes two filename arguments, a file name containing a list of lexical entries and an output file where the compiled lexicon will be saved. Compilation can take some time and may require lots of memory, as all entries are loaded in, checked and then sorted before being written out again. During compilation if some entry is malformed the reading process halts with a not so useful message. Note that if any of your entries include quote or double quotes the entries will probably be misparsed and cause such a weird error. In such cases try setting (debug_output t) before compilation. This will print out each entry as it is read in which should help to narrow down where the error is. File: festival.info, Node: Lookup process, Next: Letter to sound rules, Prev: Defining lexicons, Up: Lexicons Lookup process ============== When looking up a word, either through the C++ interface, or Lisp interface, a word is identified by its headword and part of speech. If no part of speech is specified, `nil' is assumed which matches any part of speech tag. The lexicon look up process first checks the addenda, if there is a full match (head word plus part of speech) it is returned. If there is an addenda entry whose head word matches and whose part of speech is `nil' that entry is returned. If no match is found in the addenda, the compiled lexicon, if present, is checked. Again a match is when both head word and part of speech tag match, or either the word being searched for has a part of speech `nil' or an entry has its tag as `nil'. Unlike the addenda, if no full head word and part of speech tag match is found, the first word in the lexicon whose head word matches is returned. The rationale is that the letter to sound rules (the next defence) are unlikely to be better than an given alternate pronunciation for a the word but different part of speech. Even more so given that as there is an entry with the head word but a different part of speech this word may have an unusual pronunciation that the letter to sound rules will have no chance in producing. Finally if the word is not found in the compiled lexicon it is passed to whatever method is defined for unknown words. This is most likely a letter to sound module. *Note Letter to sound rules::. Optional pre- and post-lookup hooks can be specified for a lexicon. As a single (or list of) Lisp functions. The pre-hooks will be called with two arguments (word and features) and should return a pair (word and features). The post-hooks will be given a lexical entry and should return a lexical entry. The pre- and post-hooks do nothing by default. Compiled lexicons may be created from lists of lexical entries. A compiled lexicon is _much_ more efficient for look up than the addenda. Compiled lexicons use a binary search method while the addenda is searched linearly. Also it would take a prohibitively long time to load in a typical full lexicon as an addenda. If you have more than a few hundred entries in your addenda you should seriously consider adding them to your compiled lexicon. Because many publicly available lexicons do not have syllable markings for entries the compilation method supports automatic syllabification. Thus for lexicon entries for compilation, two forms for the pronunciation field are supported: the standard full syllabified and stressed form and a simpler linear form found in at least the BEEP and CMU lexicons. If the pronunciation field is a flat atomic list it is assumed syllabification is required. Syllabification is done by finding the minimum sonorant position between vowels. It is not guaranteed to be accurate but does give a solution that is sufficient for many purposes. A little work would probably improve this significantly. Of course syllabification requires the entry's phones to be in the current phone set. The sonorant values are calculated from the _vc_, _ctype_, and _cvox_ features for the current phoneset. See `src/arch/festival/Phone.cc:ph_sonority()' for actual definition. Additionally in this flat structure vowels (atoms starting with a, e, i, o or u) may have 1 2 or 0 appended marking stress. This is again following the form found in the BEEP and CMU lexicons. Some example entries in the flat form (taken from BEEP) are ("table" nil (t ei1 b l)) ("suspicious" nil (s @ s p i1 sh @ s)) Also if syllabification is required there is an opportunity to run a set of "letter-to-sound"-rules on the input (actually an arbitrary re-write rule system). If the variable `lex_lts_set' is set, the lts ruleset of that name is applied to the flat input before syllabification. This allows simple predictable changes such as conversion of final r into longer vowel for English RP from American labelled lexicons. A list of all matching entries in the addenda and the compiled lexicon may be found by the function `lex.lookup_all'. This function takes a word and returns all matching entries irrespective of part of speech. You can optionall intercept the words as they are lookup up, and after they have been found through `pre_hooks' and `post_hooks' for each lexicon. This allows a function or list of functions to be applied to an word and feature sbefore lookup or to the resulting entry after lookup. The following example shows how to add voice specific entries to a general lexicon without affecting other voices that use that lexicon. For example suppose we were trying to use a Scottish English voice with the US English (cmu) lexicon. A number of entgries will be inapporpriate but we can redefine some entries thus (set! cmu_us_awb::lexicon_addenda '( ("edinburgh" n (((eh d) 1) ((ax n) 0) ((b r ax) 0))) ("poem" n (((p ow) 1) ((y ax m) 0))) ("usual" n (((y uw) 1) ((zh ax l) 0))) ("air" n (((ey r) 1))) ("hair" n (((hh ey r) 1))) ("fair" n (((f ey r) 1))) ("chair" n (((ch ey r) 1))))) We can the define a function that chesk to see if the word looked up is in the speaker specific exception list and use that entry instead. (define (cmu_us_awb::cmu_lookup_post entry) "(cmu_us_awb::cmu_lookup_post entry) Speaker specific lexicon addeda." (let ((ne (assoc_string (car entry) cmu_us_awb::lexicon_addenda))) (if ne ne entry))) And then for the particualr voice set up we need to add both a selection part _and_ a reset part. Thuis following the FestVox vonventions for voice set up. (define (cmu_us_awb::select_lexicon) ... (lex.select "cmu") ;; Get old var for reset and to append our function to is (set! cmu_us_awb::old_cmu_post_hooks (lex.set.post_hooks nil)) (lex.set.post_hooks (append cmu_us_awb::old_cmu_post_hooks (list cmu_us_awb::cmu_lookup_post))) ... ) ... (define (cmu_us_awb::reset_lexicon) ... ;; reset CMU's post_hooks back to original (lex.set.post_hooks cmu_us_awb::old_cmu_post_hooks) ... ) The above isn't the most efficient way as the word is looked up first then it is checked with the speaker specific list. The `pre_hooks' function are called with two arguments, the word and features, they should return a pair of word and features. File: festival.info, Node: Letter to sound rules, Next: Building letter to sound rules, Prev: Lookup process, Up: Lexicons Letter to sound rules ===================== Each lexicon may define what action to take when a word cannot be found in the addenda or the compiled lexicon. There are a number of options which will hopefully be added to as more general letter to sound rule systems are added. The method is set by the command (lex.set.lts.method METHOD) Where METHOD can be any of the following `Error' Throw an error when an unknown word is found (default). `lts_rules' Use externally specified set of letter to sound rules (described below). The name of the rule set to use is defined with the `lex.lts.ruleset' function. This method runs one set of rules on an exploded form of the word and assumes the rules return a list of phonemes (in the appropriate set). If multiple instances of rules are required use the `function' method described next. `none' This returns an entry with a `nil' pronunciation field. This will only be valid in very special circumstances. `FUNCTIONNAME' Call this as a LISP function function name. This function is given two arguments: the word and the part of speech. It should return a valid lexical entry. The basic letter to sound rule system is very simple but is powerful enough to build reasonably complex letter to sound rules. Although we've found trained LTS rules better than hand written ones (for complex languages) where no data is available and rules must be hand written the following rule formalism is much easier to use than that generated by the LTS training system (described in the next section). The basic form of a rule is as follows ( LEFTCONTEXT [ ITEMS ] RIGHTCONTEXT = NEWITEMS ) This interpretation is that if ITEMS appear in the specified right and left context then the output string is to contain NEWITEMS. Any of LEFTCONTEXT, RIGHTCONTEXT or NEWITEMS may be empty. Note that NEWITEMS is written to a different "tape" and hence cannot feed further rules (within this ruleset). An example is ( # [ c h ] C = k ) The special character `#' denotes a word boundary, and the symbol `C' denotes the set of all consonants, sets are declared before rules. This rule states that a `ch' at the start of a word followed by a consonant is to be rendered as the `k' phoneme. Symbols in contexts may be followed by the symbol `*' for zero or more occurrences, or `+' for one or more occurrences. The symbols in the rules are treated as set names if they are declared as such or as symbols in the input/output alphabets. The symbols may be more than one character long and the names are case sensitive. The rules are tried in order until one matches the first (or more) symbol of the tape. The rule is applied adding the right hand side to the output tape. The rules are again applied from the start of the list of rules. The function used to apply a set of rules if given an atom will explode it into a list of single characters, while if given a list will use it as is. This reflects the common usage of wishing to re-write the individual letters in a word to phonemes but without excluding the possibility of using the system for more complex manipulations, such as multi-pass LTS systems and phoneme conversion. From lisp there are three basic access functions, there are corresponding functions in the C/C++ domain. `(lts.ruleset NAME SETS RULES)' Define a new set of lts rules. Where `NAME' is the name for this rule, SETS is a list of set definitions of the form `(SETNAME e0 e1 ...)' and `RULES' are a list of rules as described above. `(lts.apply WORD RULESETNAME)' Apply the set of rules named `RULESETNAME' to `WORD'. If `WORD' is a symbol it is exploded into a list of the individual characters in its print name. If `WORD' is a list it is used as is. If the rules cannot be successfully applied an error is given. The result of (successful) application is returned in a list. `(lts.check_alpha WORD RULESETNAME)' The symbols in `WORD' are checked against the input alphabet of the rules named `RULESETNAME'. If they are all contained in that alphabet `t' is returned, else `nil'. Note this does not necessarily mean the rules will successfully apply (contexts may restrict the application of the rules), but it allows general checking like numerals, punctuation etc, allowing application of appropriate rule sets. The letter to sound rule system may be used directly from Lisp and can easily be used to do relatively complex operations for analyzing words without requiring modification of the C/C++ system. For example the Welsh letter to sound rule system consists or three rule sets, first to explicitly identify epenthesis, then identify stressed vowels, and finally rewrite this augmented letter string to phonemes. This is achieved by the following function (define (welsh_lts word features) (let (epen str wel) (set! epen (lts.apply (downcase word) 'newepen)) (set! str (lts.apply epen 'newwelstr)) (set! wel (lts.apply str 'newwel)) (list word nil (lex.syllabify.phstress wel)))) The LTS method for the Welsh lexicon is set to `welsh_lts', so this function is called when a word is not found in the lexicon. The above function first downcases the word and then applies the rulesets in turn, finally calling the syllabification process and returns a constructed lexically entry. File: festival.info, Node: Building letter to sound rules, Next: Lexicon requirements, Prev: Letter to sound rules, Up: Lexicons Building letter to sound rules ============================== As writing letter to sound rules by hand is hard and very time consuming, an alternative method is also available where a latter to sound system may be built from a lexicon of the language. This technique has successfully been used from English (British and American), French and German. The difficulty and appropriateness of using letter to sound rules is very language dependent, The following outlines the processes involved in building a letter to sound model for a language given a large lexicon of pronunciations. This technique is likely to work for most European languages (including Russian) but doesn't seem particularly suitable for very language alphabet languages like Japanese and Chinese. The process described here is not (yet) fully automatic but the hand intervention required is small and may easily be done even by people with only a very little knowledge of the language being dealt with. The process involves the following steps * Pre-processing lexicon into suitable training set * Defining the set of allowable pairing of letters to phones. (We intend to do this fully automatically in future versions). * Constructing the probabilities of each letter/phone pair. * Aligning letters to an equal set of phones/_epsilons_. * Extracting the data by letter suitable for training. * Building CART models for predicting phone from letters (and context). * Building additional lexical stress assignment model (if necessary). All except the first two stages of this are fully automatic. Before building a model its wise to think a little about what you want it to do. Ideally the model is an auxiluary to the lexicon so only words not found in the lexicon will require use of the letter to sound rules. Thus only unusual forms are likely to require the rules. More precisely the most common words, often having the most non-standard pronunciations, should probably be explicitly listed always. It is possible to reduce the size of the lexicon (sometimes drastically) by removing all entries that the training LTS model correctly predicts. Before starting it is wise to consider removing some entries from the lexicon before training, I typically will remove words under 4 letters and if part of speech information is available I remove all function words, ideally only training from nouns verbs and adjectives as these are the most likely forms to be unknown in text. It is useful to have morphologically inflected and derived forms in the training set as it is often such variant forms that not found in the lexicon even though their root morpheme is. Note that in many forms of text, proper names are the most common form of unknown word and even the technique presented here may not adequately cater for that form of unknown words (especially if they unknown words are non-native names). This is all stating that this may or may not be appropriate for your task but the rules generated by this learning process have in the examples we've done been much better than what we could produce by hand writing rules of the form described in the previous section. First preprocess the lexicon into a file of lexical entries to be used for training, removing functions words and changing the head words to all lower case (may be language dependent). The entries should be of the form used for input for Festival's lexicon compilation. Specifical the pronunciations should be simple lists of phones (no syllabification). Depending on the language, you may wish to remve the stressing--for examples here we have though later tests suggest that we should keep it in even for English. Thus the training set should look something like ("table" nil (t ei b l)) ("suspicious" nil (s @ s p i sh @ s)) It is best to split the data into a training set and a test set if you wish to know how well your training has worked. In our tests we remove every tenth entry and put it in a test set. Note this will mean our test results are probably better than if we removed say the last ten in every hundred. The second stage is to define the set of allowable letter to phone mappings irrespective of context. This can sometimes be initially done by hand then checked against the training set. Initially constract a file of the form (require 'lts_build) (set! allowables '((a _epsilon_) (b _epsilon_) (c _epsilon_) ... (y _epsilon_) (z _epsilon_) (# #))) All letters that appear in the alphabet should (at least) map to `_epsilon_', including any accented characters that appear in that language. Note the last two hashes. These are used by to denote beginning and end of word and are automatically added during training, they must appear in the list and should only map to themselves. To incrementally add to this allowable list run festival as festival allowables.scm and at the prompt type festival> (cummulate-pairs "oald.train") with your train file. This will print out each lexical entry that couldn't be aligned with the current set of allowables. At the start this will be every entry. Looking at these entries add to the allowables to make alignment work. For example if the following word fails ("abate" nil (ah b ey t)) Add `ah' to the allowables for letter `a', `b' to `b', `ey' to `a' and `t' to letter `t'. After doing that restart festival and call `cummulate-pairs' again. Incrementally add to the allowable pairs until the number of failures becomes accceptable. Often there are entries for which there is no real relationship between the letters and the pronunciation such as in abbreviations and foreign words (e.g. "aaa" as "t r ih p ax l ey"). For the lexicons I've used the technique on less than 10 per thousand fail in this way. It is worth while being consistent on defining your set of allowables. (At least) two mappings are possible for the letter sequence `ch'--having letter `c' go to phone `ch' and letter `h' go to `_epsilon_' and also letter `c' go to phone `_epsilon_' and letter `h' goes to `ch'. However only one should be allowed, we preferred `c' to `ch'. It may also be the case that some letters give rise to more than one phone. For example the letter `x' in English is often pronunced as the phone combination `k' and `s'. To allow this, use the multiphone `k-s'. Thus the multiphone `k-s' will be predicted for `x' in some context and the model will separate it into two phones while it also ignoring any predicted `_epsilons_'. Note that multiphone units are relatively rare but do occur. In English, letter `x' give rise to a few, `k-s' in `taxi', `g-s' in `example', and sometimes `g-zh' and `k-sh' in `luxury'. Others are `w-ah' in `one', `t-s' in `pizza', `y-uw' in `new' (British), `ah-m' in `-ism' etc. Three phone multiphone are much rarer but may exist, they are not supported by this code as is, but such entries should probably be ignored. Note the `-' sign in the multiphone examples is significant and is used to indentify multiphones. The allowables for OALD end up being (set! allowables ' ((a _epsilon_ ei aa a e@ @ oo au o i ou ai uh e) (b _epsilon_ b ) (c _epsilon_ k s ch sh @-k s t-s) (d _epsilon_ d dh t jh) (e _epsilon_ @ ii e e@ i @@ i@ uu y-uu ou ei aa oi y y-u@ o) (f _epsilon_ f v ) (g _epsilon_ g jh zh th f ng k t) (h _epsilon_ h @ ) (i _epsilon_ i@ i @ ii ai @@ y ai-@ aa a) (j _epsilon_ h zh jh i y ) (k _epsilon_ k ch ) (l _epsilon_ l @-l l-l) (m _epsilon_ m @-m n) (n _epsilon_ n ng n-y ) (o _epsilon_ @ ou o oo uu u au oi i @@ e uh w u@ w-uh y-@) (p _epsilon_ f p v ) (q _epsilon_ k ) (r _epsilon_ r @@ @-r) (s _epsilon_ z s sh zh ) (t _epsilon_ t th sh dh ch d ) (u _epsilon_ uu @ w @@ u uh y-uu u@ y-u@ y-u i y-uh y-@ e) (v _epsilon_ v f ) (w _epsilon_ w uu v f u) (x _epsilon_ k-s g-z sh z k-sh z g-zh ) (y _epsilon_ i ii i@ ai uh y @ ai-@) (z _epsilon_ z t-s s zh ) (# #) )) Note this is an exhaustive list and (deliberately) says nothing about the contexts or frequency that these letter to phone pairs appear. That information will be generated automatically from the training set. Once the number of failed matches is signficantly low enough let `cummulate-pairs' run to completion. This counts the number of times each letter/phone pair occurs in allowable alignments. Next call festival> (save-table "oald-") with the name of your lexicon. This changes the cummulation table into probabilities and saves it. Restart festival loading this new table festival allowables.scm oald-pl-table.scm Now each word can be aligned to an equally-lengthed string of phones, epsilon and multiphones. festival> (aligndata "oald.train" "oald.train.align") Do this also for you test set. This will produce entries like aaronson _epsilon_ aa r ah n s ah n abandon ah b ae n d ah n abate ah b ey t _epsilon_ abbe ae b _epsilon_ iy The next stage is to build features suitable for `wagon' to build models. This is done by festival> (build-feat-file "oald.train.align" "oald.train.feats") Again the same for the test set. Now you need to constructrure a description file for `wagon' for the given data. The can be done using the script `make_wgn_desc' provided with the speech tools Here is an example script for building the models, you will need to modify it for your particualr database but it shows the basic processes for i in a b c d e f g h i j k l m n o p q r s t u v w x y z do # Stop value for wagon STOP=2 echo letter $i STOP $STOP # Find training set for letter $i cat oald.train.feats | awk '{if ($6 == "'$i'") print $0}' >ltsdataTRAIN.$i.feats # split training set to get heldout data for stepwise testing traintest ltsdataTRAIN.$i.feats # Extract test data for letter $i cat oald.test.feats | awk '{if ($6 == "'$i'") print $0}' >ltsdataTEST.$i.feats # run wagon to predict model wagon -data ltsdataTRAIN.$i.feats.train -test ltsdataTRAIN.$i.feats.test \ -stepwise -desc ltsOALD.desc -stop $STOP -output lts.$i.tree # Test the resulting tree against wagon_test -heap 2000000 -data ltsdataTEST.$i.feats -desc ltsOALD.desc \ -tree lts.$i.tree done The script `traintest' splits the given file `X' into `X.train' and `X.test' with every tenth line in `X.test' and the rest in `X.train'. This script can take a significnat amount of time to run, about 6 hours on a Sun Ultra 140. Once the models are created the must be collected together into a single list structure. The trees generated by `wagon' contain fully probability distributions at each leaf, at this time this information can be removed as only the most probable will actually be predicted. This substantially reduces the size of the tress. (merge_models 'oald_lts_rules "oald_lts_rules.scm") (`merge_models' is defined within `lts_build.scm') The given file will contain a `set!' for the given variable name to an assoc list of letter to trained tree. Note the above function naively assumes that the letters in the alphabet are the 26 lower case letters of the English alphabet, you will need to edit this adding accented letters if required. Note that adding "'" (single quote) as a letter is a little tricky in scheme but can be done--the command `(intern "'")' will give you the symbol for single quote. To test a set of lts models load the saved model and call the following function with the test align file festival oald-table.scm oald_lts_rules.scm festival> (lts_testset "oald.test.align" oald_lts_rules) The result (after showing all the failed ones), will be a table showing the results for each letter, for all letters and for complete words. The failed entries may give some notion of how good or bad the result is, sometimes it will be simple vowel diferences, long versus short, schwa versus full vowel, other times it may be who consonants missing. Remember the ultimate quality of the letter sound rules is how adequate they are at providing _acceptable_ pronunciations rather than how good the numeric score is. For some languages (e.g. English) it is necessary to also find a stree pattern for unknown words. Ultimately for this to work well you need to know the morphological decomposition of the word. At present we provide a CART trained system to predict stress patterns for English. If does get 94.6% correct for an unseen test set but that isn't really very good. Later tests suggest that predicting stressed and unstressed phones directly is actually better for getting whole words correct even though the models do slightly worse on a per phone basis `black98'. As the lexicon may be a large part of the system we have also experimented with removing entries from the lexicon if the letter to sound rules system (and stree assignment system) can correct predict them. For OALD this allows us to half the size of the lexicon, it could possibly allow more if a certain amount of fuzzy acceptance was allowed (e.g. with schwa). For other languages the gain here can be very signifcant, for German and French we can reduce the lexicon by over 90%. The function `reduce_lexicon' in `festival/lib/lts_build.scm' was used to do this. A diccussion of using the above technique as a dictionary compression method is discussed in `pagel98'. A morphological decomposition algorithm, like that described in `black91', may even help more. The technique described in this section and its relative merits with respect to a number of languages/lexicons and tasks is dicussed more fully in `black98'. File: festival.info, Node: Lexicon requirements, Next: Available lexicons, Prev: Building letter to sound rules, Up: Lexicons Lexicon requirements ==================== For English there are a number of assumptions made about the lexicon which are worthy of explicit mention. If you are basically going to use the existing token rules you should try to include at least the following in any lexicon that is to work with them. * The letters of the alphabet, when a token is identified as an acronym it is spelled out. The tokenization assumes that the individual letters of the alphabet are in the lexicon with their pronunciations. They should be identified as nouns. (This is to distinguish `a' as a determiner which can be schwa'd from `a' as a letter which cannot.) The part of speech should be `nn' by default, but the value of the variable `token.letter_pos' is used and may be changed if this is not what is required. * One character symbols such as dollar, at-sign, percent etc. Its difficult to get a complete list and to know what the pronunciation of some of these are (e.g hash or pound sign). But the letter to sound rules cannot deal with them so they need to be explicitly listed. See the list in the function `mrpa_addend' in `festival/lib/dicts/oald/oaldlex.scm'. This list should also contain the control characters and eight bit characters. * The possessive `'s' should be in your lexicon as schwa and voiced fricative (`z'). It should be in twice, once as part speech type `pos' and once as `n' (used in plurals of numbers acronyms etc. e.g 1950's). `'s' is treated as a word and is separated from the tokens it appears with. The post-lexical rule (the function `postlex_apos_s_check') will delete the schwa and devoice the `z' in appropriate contexts. Note this post-lexical rule brazenly assumes that the unvoiced fricative in the phoneset is `s'. If it is not in your phoneset copy the function (it is in `festival/lib/postlex.scm') and change it for your phoneset and use your version as a post-lexical rule. * Numbers as digits (e.g. "1", "2", "34", etc.) should normally _not_ be in the lexicon. The number conversion routines convert numbers to words (i.e. "one", "two", "thirty four", etc.). * The word "unknown" or whatever is in the variable `token.unknown_word_name'. This is used in a few obscure cases when there just isn't anything that can be said (e.g. single characters which aren't in the lexicon). Some people have suggested it should be possible to make this a sound rather than a word. I agree, but Festival doesn't support that yet. File: festival.info, Node: Available lexicons, Next: Post-lexical rules, Prev: Lexicon requirements, Up: Lexicons Available lexicons ================== Currently Festival supports a number of different lexicons. They are all defined in the file `lib/lexicons.scm' each with a number of common extra words added to their addendas. They are `CUVOALD' The Computer Users Version of Oxford Advanced Learner's Dictionary is available from the Oxford Text Archive `ftp://ota.ox.ac.uk/pub/ota/public/dicts/710'. It contains about 70,000 entries and is a part of the BEEP lexicon. It is more consistent in its marking of stress though its syllable marking is not what works best for our synthesis methods. Many syllabic `l''s, `n''s, and `m''s, mess up the syllabification algorithm, making results sometimes appear over reduced. It is however our current default lexicon. It is also the only lexicon with part of speech tags that can be distributed (for non-commercial use). `CMU' This is automatically constructed from `cmu_dict-0.4' available from many places on the net (see `comp.speech' archives). It is not in the mrpa phone set because it is American English pronunciation. Although mappings exist between its phoneset (`darpa') and `mrpa' the results for British English speakers are not very good. However this is probably the biggest, most carefully specified lexicon available. It contains just under 100,000 entries. Our distribution has been modified to include part of speech tags on words we know to be homographs. `mrpa' A version of the CSTR lexicon which has been floating about for years. It contains about 25,000 entries. A new updated free version of this is due to be released soon. `BEEP' A British English rival for the `cmu_lex'. BEEP has been made available by Tony Robinson at Cambridge and is available in many archives. It contains 163,000 entries and has been converted to the `mrpa' phoneset (which was a trivial mapping). Although large, it suffers from a certain randomness in its stress markings, making use of it for synthesis dubious. All of the above lexicons have some distribution restrictions (though mostly pretty light), but as they are mostly freely available we provide programs that can convert the originals into Festival's format. The MOBY lexicon has recently been released into the public domain and will be converted into our format soon. File: festival.info, Node: Post-lexical rules, Prev: Available lexicons, Up: Lexicons Post-lexical rules ================== It is the lexicon's job to produce a pronunciation of a given word. However in most languages the most natural pronunciation of a word cannot be found in isolation from the context in which it is to be spoken. This includes such phenomena as reduction, phrase final devoicing and r-insertion. In Festival this is done by post-lexical rules. `PostLex' is a module which is run after accent assignment but before duration and F0 generation. This is because knowledge of accent position is necessary for vowel reduction and other post lexical phenomena and changing the segmental items will affect durations. The `PostLex' first applies a set of built in rules (which could be done in Scheme but for historical reasons are still in C++). It then applies the functions set in the hook `postlex_rules_hook'. These should be a set of functions that take an utterance and apply appropriate rules. This should be set up on a per voice basis. Although a rule system could be devised for post-lexical sound rules it is unclear what the scope of them should be, so we have left it completely open. Our vowel reduction model uses a CART decision tree to predict which syllables should be reduced, while the "'s" rule is very simple (shown in `festival/lib/postlex.scm'). The `'s' in English may be pronounced in a number of different ways depending on the preceding context. If the preceding consonant is a fricative or affricative and not a palatal labio-dental or dental a schwa is required (e.g. `bench's') otherwise no schwa is required (e.g. `John's'). Also if the previous phoneme is unvoiced the "s" is rendered as an "s" while in all other cases it is rendered as a "z". For our English voices we have a lexical entry for "'s" as a schwa followed by a "z". We use a post lexical rule function called `postlex_apos_s_check' to modify the basic given form when required. After lexical lookup the segment relation contains the concatenation of segments directly from lookup in the lexicon. Post lexical rules are applied after that. In the following rule we check each segment to see if it is part of a word labelled "'s", if so we check to see if are we currently looking at the schwa or the z part, and test if modification is required (define (postlex_apos_s_check utt) "(postlex_apos_s_check UTT) Deal with possesive s for English (American and British). Delete schwa of 's if previous is not a fricative or affricative, and change voiced to unvoiced s if previous is not voiced." (mapcar (lambda (seg) (if (string-equal "'s" (item.feat seg "R:SylStructure.parent.parent.name")) (if (string-equal "a" (item.feat seg 'ph_vlng)) (if (and (member_string (item.feat seg 'p.ph_ctype) '(f a)) (not (member_string (item.feat seg "p.ph_cplace") '(d b g)))) t;; don't delete schwa (item.delete seg)) (if (string-equal "-" (item.feat seg "p.ph_cvox")) (item.set_name seg "s")))));; from "z" (utt.relation.items utt 'Segment)) utt) File: festival.info, Node: Utterances, Next: Text analysis, Prev: Lexicons, Up: Top Utterances ********** The utterance structure lies at the heart of Festival. This chapter describes its basic form and the functions available to manipulate it. * Menu: * Utterance structure:: internal structure of utterances * Utterance types:: Type defined synthesis actions * Example utterance types:: Some example utterances * Utterance modules:: * Accessing an utterance:: getting the data from the structure * Features:: Features and features names * Utterance I/O:: Saving and loading utterances