Paper



i. introduction
broad phonetic technique -emphasize what is reliable, instead of everything probabilistic at the onset
     advantage: do't have to set probabilities
     disadvantage:  there may be information there that would obviate need for cohort construction
Question:  what kind of questions can we ask to reveal what this technique is best suited for, what are any inherent limitations?  How can it be augented, modified, or what other algorithms are appropriate for other domains?


I.  Broad phonetic classification

ZSH Only linear precedence information
strings of segments 
no constraints
no suprasegmental information
notion of equivalence class; tailored for isolated word recognition
anything like it before??
  Waltz?  Delayed binding?


II.  A different dimension is stress

DPH How it cuts across equivalence classes
as reliable as BP categorization
Aull

III. What about continuous speech with BPC with or without stress?

There are only around 250 BP syllables (??times 2 with stress)
(reason 6k allophonic ones (fujimura & Lovins)  this equals 250 BP ones, cut out vowels, collapse stops in 3? positions; etc.)

SO combinatorics 
large cohort sizes with big dictionaries

KWC on this

IV. Possibly ok with bounds on lexicon size for continuous

V.  Need other source of constraint for big word lists, continuous case
(based on III and IV)


one way is longer units than syllables, so more than 250 divisions

VI.  Feet is a suprasyllabic unit that incorporates stress, has constraint

-- better equivalence class size (experiments)
-- limited problem with boundaries

VII.  What about deletion? Unreliability?

Don't specify all BPC equally.  Specify Stress in more fine-grain detail, others as empty syllables.



VIII.  Back to deletion.  There are systematicities in deletions, and constraints on strings across syllables

Want constraints on strings across syllables because can get huge cohorts otherwise since word boundaries are unknown

Assume: Vowel in primary stress in word will not delete if content word

"Coming at you fast" (withgott 85)

No point in gathering the cohort for certain patterns:

no *tyu - f V:
no *Cx - f V: 
no  *tx - f V:
no  *t/C x - Fric V:
no  *t/CFric V: (one syllable)
no  *?ty x- Fric V:
no  *y x- Fric V:
no  *?ty x- Fric V: (one syllable) 
no  *y x- Fric V: (one syllable) 

No point in gathering the cohort for (?)(t)y (x) FRIC Stress V

cuts back on combinatorics, even with function words



IX. lexical vs. metrical


we are not dealing with the theorists classic feet here, because "about"

X. Stress in the signal

.. how reliable?

what about stress shift?
This is not so bad because easily detectable stress (by current algorithms) is only primary, unreduced and reduced  anyway,  so let us only rely on it at the onset

artificial intelligence -->
ARti'ficial inTELligence

look up all artx all fISl   -- art won't be in lexicon  as reduced xrtx

fISl wont reduce all the way down,  so tractable



XI.  Conclusion