CS 551651:

1
CS 551/651 Structure of Spoken Language Lecture
7 Syllable Structure, Vowel Neutralization, and
Coarticulation John-Paul Hosom Fall 2008
2

• NOTE
• Theres a tutorial on the web that allows you
  tohear the effect of different formant values
• http//www.asel.udel.edu/speech/tutorials/synthesi
  s/ceevees.html
• You can enter start time, end time, amplitude,
  and formant
• values for beginning, middle and end of a
  syllable, then
• generate a waveform and hear the result.

3

• Syllables
• Words are composed of phonetic clusters
  syllables
• Each syllable has a nucleus typically the
  nucleus isa vowel or diphthong, sometimes a
  syllabic nasal or lateral (button, bottle) or
  retroflex (bird)
• Nucleus is syllabic nasal or lateral only when
  following alveolar consonant in previous
  syllable of a word
• Syllable boundaries sometimes ambiguous tasty
  tas/ty tast/y ta/sty bottling bott/l/ing
  bott/ling
• Syllable can be broken into components syllable
  contains onset, rhyme rhyme contains
  nucleus, codaonset and coda are consonants,
  rhyme is a vowel.

4
Syllables
Limitations on consonant clusters not all CCC
combinations are possible in syllable-initial
position. Of those that are possible, almost
half are very rare.
possibly only one word in English spew
only a few English words pronounced (optionally)
with /s t y/ Stewart, steward, stew
very few English words/root with /s k l/
sclerosis
very few English words with /s k y/ skew,
askew, obscure
graphic from http//www.arts.uwa.edu.au/LingWWW/LI
N101-102
5

• Syllables
• Sonority corresponds roughly to degree of
  constrictionalong vocal and/or nasal tract
• Ordering of sonority vowels, glides (/w/,
  /y/), liquids (/l/, /r/), nasals, fricatives,
  affricates, plosives
• If a binary classification (sonorant/non-sonorant)
  , then sonorant consists of all vowels, glides,
  liquids, and nasals.
• Fricatives, affricates, and plosives may be
  clustered into onecategory, obstruents, for
  purposes of sonority
• Syllabification can be done according to
  sonority principlethe sonority must rise and
  fall in a syllable
• Also, theres the Maximal Onset PrinciplePut a
  consonant in the onset rather than the coda when
  possible

6

• Syllables
• Because of rise and fall of sonority in
  syllables, the followingrestrictions occur
  (a) glide (/w/,/y/) must be immediately
  adjacent to a vowel, (b) /r/ is next closest
  consonant to vowel, (c) /l/ is next closest
  consonant to vowel, (d) nasal is next
  closest, (e) obstruent is farthest from the
  vowel (but there may be more than one
  obstruent in onset or coda)
• Obstruents in a cluster must have same voicing
• In series of obstruents between two vowels,
  voicing can change only once, at the syllable
  boundary.
• English allows up to 3 consonants in syllable
  initial position, 4 consonants at syllable final
  position

7

• Syllables
• Examples sphere /s f iy r/, streak /s t r iy k/,
  texts /t eh k s t s/, helms /h eh l m z/ but
  not /t l iy/ or /p w iy/
• The ordering of glides and liquids doesnt matter
  for our purposes (applying to syllabification),
  because glides and liquids can not occur
  sequentially within the same syllablein English.
  (However, two liquids in the same syllable
  ispossible, e.g. Carl and girl, as long as
  /r/ is closer toto the vowel than /l/.)
• In English, most burst-fricative pairs are
  represented as distinct phonemes (/ch/, /jh/),
  although there are some othercases of
  burst-fricative pairs (e.g. tsunami,
  bishops).
• Its also possible to have two or more adjacent
  fricatives eleven twelfths

8

• Vowel Neutralization
• When speech is uttered very quickly (or is not
  well enunciated),the formants tend to shift
  toward that of a neutral vowel

(from Daniloff, p. 320)
(from van Bergem 1993 p. 8)
9

• Vowel Neutralization
• Target undershoot

/m ih pc ph ih
eh/
10

• Vowel Neutralization

/m ih pc ph ih
eh/
Target undershoot /ih/ extracted and
concatenated from mip
11

• Vowel Neutralization
• However, neutralization is not always so simple
  sometimesvowel formants shift away from the
  neutral position,depending on their context, and
  vowels tend toward slightlydifferent neutral
  targets.
• Neutralization is to some extent an artifact of
  averagingover speakers and contexts (van Bergem
  1993)

vowels from one speaker in different
phonetic contexts, and in reduced and
isolated speaking conditions
12

• Coarticulation
• Coarticulation is the blending of adjacent
  speech sounds,
• due to gradual movement of the articulators.
• Coarticulation makes automatic speech recognition
  andtext-to-speech synthesis difficult, but
  humans use coarticulationto conserve effort
  while speaking and provide robustnessduring
  recognition.
• There is Right-to-Left (RL) or anticipatory and
  Left-to-Right (LR) or carry-over
  coarticulation
• Models of coarticulation and syllabification
  ? Locus Theory ? Modified Locus Theory
  (Klatt) ? Öhmans Theory ? Kozhevnikov-Chistov
  ich (KC) Theory ? Wickelgrens Theory, etc.

13
Coarticulation RL coarticulation occurs
due to high-level planning of phonetic sequences
spoon s p uw n rounding in
isolation rounding in context
more observable if neighboring sounds
not specified with respect to potentially
coarticulated feature e.g. /s/, /p/, /n/ not
specified with respect to lip rounding (from
Daniloff, pp. 323-324)
14
Coarticulation Locus Theory Locus Theory
(Delattre, Liberman, and Cooper, 1955) there
are, for each consonant, characteristic frequency
positions, or loci, at which the formant
transitions begin, or to which they may be
assumed to point. On this basis, the transitions
may be regarded simply as movements of the
formants from their respective loci to the
frequency levels appropriate for the next phone
The spectrographic patterns , which produce /d/
before /iy/, /aa/, and /ow/, show how these
transitions seem to be pointing to a F2 locus
in the vicinity of 1800 Hz. ? Each consonant
has target frequencies independent of
the neighboring vowels. ? Formants transition
from these target frequencies to the
vowel target frequencies.
15

• Coarticulation Locus Theory
• Locus Theory
• Consonants and vowels both have targets of
  articulatorpositions and therefore formant
  frequency locations
• Given sufficient duration of a syllable, all
  phonemes reachtheir targets
• The slope of the formants during a transition
  from a consonantto a vowel is relatively
  constant until reaching the target
• If the syllable duration doesnt allow enough
  time for theformants to reach their targets,
  target undershoot occursand the formants
  change direction before fully realizingthe
  intended vowel

16

• Coarticulation Locus Theory
• Locus Theory

(From Klatt 1987, p. 753)
17

• Coarticulation Modified Locus Theory
• Problems with Locus Theory
• A transition may have both rapid and slow
  componentsrapid release of obstruction via
  tongue tip, followed by slow movement of tongue
  body.
• Preceding vowel can influence F2 onset of a CV
  transition(Öhman, 1966)
• F2 may be insensitive to oral constrictions
  (obstruents)if the tongue position is toward the
  front of the mouth (as in /iy/)
• (as reported by Fant 1973, Klatt1987)

18

• Coarticulation Modified Locus Theory
• Modified Locus Theory
• Klatt hypothesized that main effects of the vowel
  on thearticulation of consonants are front/back
  position and liprounding
• Vowels divided into three sets front round
  front, round(because there are no rounded
  front vowels in English,sets 1 and 2 are
  mutually exclusive)
• front /iy ih eh ae/
• round /uw ao ow er/
• front, round /uh ah aa aw/
• Predicted Fonset from Ftarget for these 3 classes
  (locus theory)
• Achieved 95 intelligibility for CVC nonsense
  syllables

19

• Coarticulation Locus Theory
• Modified Locus Theory

-front, -round front round
(From Klatt 1987, p. 754)
20
Coarticulation Öhmans Theory Öhman (1965)
found that loci of consonants is NOT
independent of neighboring vowels a
nd that for /g/ more than one locus is
required Conclusion consonant gestures are
superimposed on vowel gestures that are present
during the consonant even when consonant is
being uttered in VCV, there is effect of both V
on C.
21
Coarticulation Öhmans Theory Öhman (1966)
proposed model of coarticulation based
on vocal-tract shape evolving over time. Assumes
that vocal-tract shapes can be mapped to formant
frequencies. For VCV utterances where s(x,t)
is the vocal tract shape at position x and time
t, v(x) is the vocal tract shape at position x
for a given vowel, c(x) is the vocal tract shape
of the consonant, k(t) is an interpolation value
(from 0 to 1), and wc(x) describes the degree to
which c(x) resists coarticulation. v(x)
describes the shape of the vocal tract, which
may be a combination of two vowels if V1 ? V2.
(v(x) will vary over time from V1 to V2)
22

• Coarticulation Kozhevnikov-Chistovich (KC)
  Theory
• Syllabification using CnV pattern CV, CCV,
  CCCV,
• phrase give true answers
• g ih v t r uw ae n s er z
• ---- ----------- -- ------- -
• S1 S2 S3 S4 S5
• (2) Measured relative durations of words,
  syllables, vowels
• relative duration of vowel Dvow / Dsyll,
  syllable Dsyll / Dword word Dword /
  Dphrase

23
Coarticulation Kozhevnikov-Chistovich (KC)
Theory Found coarticulation within syllable but
not across syllables C1 V1 C2 C3 V2

• articulatory gestures for consonant(s) and vowel
  begin nearly
• simultaneously with onset of initial consonant in
  syllable
• Example lip rounding in /uw/ begins with /v/ in
  give true answers,
• but nasalization of /ae/ does not occur.
• assumes little or no LR coarticulation
• assumes motor programming of speech is
  discontinuous at VC boundary
• counter-examples showing LR coarticulation
  (Moll and Daniloff 1971, Kent, Carney, and
  Severeid 1974, Öhman 1966)

24
Coarticulation Wickelgrens Theory Speech
units are mentally coded as context-sensitive
units in phonetic string /X Y Z/, Y is encoded
as XYZ By assuming (context-sensitive)
allophones to be the basic unit of articulation,
it is trivial to account for how the
same phoneme in different phonemic environments
can be different in some respects at all levels
of the speech process (Wickelgren 1969, p. 11)
However, coarticulation can spread over more
than one phone (up to seven phones distance).
Other criticisms MacNeilage 1970, Whitaker
1970, Halwes and Jenkins 1971 Allophonic
richness may only beget strategic poverty (Kent
and Minifie 1977) However, Wickelgrens is the
only model currently used in ASR and
concatenative text-to-speech (exceptions Wouters
2001, Wrede 2001).
25

• Coarticulation Gays Theory
• Gay, 1977 The syllabic unit of motor
  organization is the CV unit
• Based on X-ray motion pictures of VCV utterances
• anticipatory tongue movements for V2 in V1CV2
  sequencedont begin until closure of C has been
  attained
• movement toward V2 occurs during closure of C,
  havinga large effect on position and shape of
  tongue during releaseof closure
• V1 has little effect on position of tongue at
  moment ofclosure
• supports KC theory conflicts with Ohmans
  findings

26

• Coarticulation
• Other models MacNeilage, Henke, Benguerel and
  Cowan,Moll and Daniloff, Liberman, Tatham, etc.
• Some are feature based in that each phonetic
  segmentis assigned distinctive features which
  can then be modifiedin regular ways
• Some are hierarchical models, with several
  levels oforganization and complex interaction
  between levels
• However, coarticulatory patterns are not
  explainedadequately by any theories or models
  (Kent and Minifie, 1977)
• Conflicting evidence (Öhman and Kent Moll vs.
  KC and Gay)