Go to the
first,
previous,
next,
last section,
table of contents.
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.
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)) )
See section 13.3 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.
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, See section 13.4 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.
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. See section 13.4 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.
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.
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.
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.
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.
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)
Go to the first, previous, next, last section, table of contents.
