Ling/CSE 472: Assignment 5: Text-to-Speech

Due November 30th, by the start of class

Text-to-speech systems rely on large pronunciation dictionaries in combination with rules for dealing with unknown words. One large class of typically unknown words is names, including people's names. This assignment will focus on one particular aspect of predicting the pronunciation of names, namely stress assignment. In particular, there are `name suffixes' (recurring forms at the end of many different names) which leave the stress assignment of the stem unchanged (stress-neutral name suffixes) and those which cause the stress to move (stress-changing suffixes). It is stress assignment is important for figuring out pronunciation because a) phonological rules affecting the pronunciation of segments are sensitive to stress and b) lexical stress interacts with other factors to determine the prosody of an utterance.

Your tasks

• Find 6 name stems (surname stems; perhaps by consulting a phone book), with the following properties:
  1. A one-syllable name stem
  2. A two-syllable name stem with stress on the first syllable
  3. A two-syllable name stem with stress on the second syllable
  4. A three-syllable name stem with stress on the first syllable
  5. A three-syllable name stem with stress on the second syllable
  6. A three-syllable name stem with stress on the third syllable
  Indicate stress by placing a ' before the vowel of the stressed syllable.
• Find two name suffixes, one which is stress neutral and one which changes the stress on the stem.
• Determine how the stress changes: write out all 12 names you can make by attaching one or the other suffix to your 6 names, and indicate in this list where the (primary) stress in each one, using the ' notation as above. This list is what you will use to evaluate whether or not your system is working.
• Write a regular expression which represents the possible underlying forms (stem only; stem + suffix, with the morpheme boundary and underlying stress indicated). You may find it useful to use the `explode' operators { }. In xfst regular expressions, characters not separated by spaces are treated as multicharacter symbols. This is not what you want. Therefore, (part of) your regular expression will either have to look like this:

  [ c a t | d o g ]
  

  Or, using the `explode' operators, like this:

  [ {cat} | {dog} ]
  

  Recall that % is the escape character, and that most punctuation marks have a special meaning in xfst and therefore need to be escaped. More on xfst syntax can be found here.

• Write a rule (or rules, if necessary) to relate the underlying forms to the surface forms. You are welcome to use the xfst rule notations for this assignment. In particular, note the following:

  [ A -> B || C _ D ]
  

  This is the xfst form for A is rewritten as B when it occurs between C and D (where C and D refer to the upper tape context). A, B, C and D are all regular expressions.

  Rules that run in parallel are separated by ,,:

  [ A -> B || C _ D ,, E -> F || G _ H ]
  

  Rules of epenthesis (which insert something where nothing was before) are written like this:

  [ [..] -> A || B _ C ]
  

  This can be read as "nothing goes to A in the context B _ C". If you used 0 (epsilon) instead of [..] in this rule, it would try to insert an infinite number of As between B and C, because there are an infinite number of empty strings between B and C.
• Write a rule which removes the morpheme boundary, so that it does not appear in the surface forms.
• Combine the above into an xfst script. The script should create one FST composed out of the underlying forms regular expression and the rules. (For reference, you might want to look at k.xfst from Assignment 2.)
• Test your FSTs by loading them and then calling "print lower". This will print the strings of the lower-tape language. (You can use "print upper" to see the upper tape language.) You may find it convenient to put this command at the end of your script.
• Describe (in prose) three phonological changes that you observe in your data which correlate with stress. (I.e., in what ways do the pronunciations of the name stems change when the stress is moved?)

Turn in

• Six name stems, with stress indicated.
• Two name suffixes.
• Twelve stem+suffix names, with stress indicated.
• Your XFST script containing the FST which accepts/generates all 18 names with the right (primary) stress assignment.
• Your answer to the question about effects of stress.

Don't worry about

• Modeling aspects of pronunciation other than stress in your xfst script. Represent the names with their orthographies (i.e., not phonetic representations).
• Which name suffixes usually appear with which names. Pretend for the purposes of this assignment that any suffix can go with any name.
• Secondary stress. Just keep track of the primary stress on each name.