CS 551/651:

1
CS 551/651 Structure of Spoken Language Lecture
6 Phonological Processes John-Paul Hosom Fall
2008
2

• Phonological Processes
• Phonemes undergo systematic variation depending
  on theircontext
• For example, forming the past tense cause /k
  aa z/ ? caused /k aa z d/ talk /t aa k/ ?
  talked /t aa k t//d/ vs. /t/ is predictable
  based on voicing of word-final phoneme
• Allophones can be viewed as systematic variations
  of phonemesthat are a result of cultural and
  physiological processes, butdo not distinguish
  meaning of utterance
• For example, /p/ and /ph/ in English is
  predictable word or syllable initial voiceless
  stops are aspiratedpit ? ph ih th tip ? th
  ih ph kin ? kh ih nspit ? s p ih
  th stick ? s t ih kh skin ? s k ih n

3
Phonological Processes
/ph ih th th ih
ph kh ih n/
/s p ih th s t
ih kh s k ih n/
4

• Phonological Processes
• Other types of phonetic processesAssimilation,
  Deletion, Reduction, Insertion,
  Substitution,Me'tathesis (switching order of two
  phonemes)
• AssimilationA feature of one segment is shared
  by a neighboring segment
• Examples of Assimilation? Nasalization of vowels
  before nasal consonants? in- (negative prefix)
  becomes im- in words beginning with bilabial
  consonant (imbalance, imperfect, indifferent,
  intolerance)

5

• Phonological Processes
• Assimilation may be due to coarticulation, or it
  may belanguage-specific, arbitrary word-fi
  nal alveolar obstruent may take on place of
  articulation of following word-initial
  segment if word-initial segment is
  palato-alveoarthis /dh ih s/ shop /sh aa
  ph/ ? this shop /dh ih sh sh aa ph/this /dh ih
  s/ fish /f ih sh/ ? this fish /dh ih s f ih
  sh/this /dh ih s/ thing /th ih ng/ ? this thing
  /dh ih s th ih ng/
• also, depending on dialect, not within-word
• misshapen /m ih s sh ei p en/

6

• Phonological Processes
• Example of assimilation of /s/ with /sh/ but not
  /f/

/dh ih sh sh aa pcl ph dh ih s f
ih sh/
7

• Phonological Processes
• Substitution common in foreign accents or
  speaking impairments welcome /v eh l k ah
  m/ McDonald /m a k uw d ow n aa r uw d
  ow/ Roger /w aa jh er/
• Metathesis changing order of two phonemes
  within a word (dialect variation) pretty /p
  er dx iy/ ask /ae k s/

8

• Phonological Processes
• Deletion Barbara /b aa r b ax r ah/ ? /b aa
  r b r ah/ Memory /m eh m ax r iy/ ? /m eh m r
  iy/
• Reduction unstressed vowels become
  /ax/ conduct (verb) /k ax n d ah k
  t/ conduct (noun) /k aa n d ax k t/
• Insertion voiceless stop inserted between
  nasal and voiceless consonant voiceless stop
  always has same place of articulation as nasal
  fancy /f ae n t s iy/ Chomsky /ch aa m p
  s k iy/ schwa inserted after word-final
  nasal nine /n ay n ax/

dictionary pronunciation
9

• Phonological Processes
• Deletion

/m eh m r
iy/
10

• Phonological Processes
• Insertion

/f ae n t s iy ch
aa m p s k iy/
11

• Phonological Processes Ladefoged Rules
• voiced, stop ? aspirated when syllable
  initial pit vs. spit
• ax ? voiced after syllable-initial voiced,
  stop and before voiced, stop potato
• consonantal ? longer at end of phrase bib,
  did, don, nod
• voiced, stop ? aspirated after
  syllable-initial /s/ spew, stew, skew
• vowel ? shorter before unvoiced phonemes in
  same syllable cap vs. cab, back vs. bag

12

• Phonological Processes Ladefoged Rules
• Devoicing, End-of-Phrase Length

/ph ax tcl th ey dx
ow/
/d aa n
n aa dcl d/
13

• Phonological Processes Ladefoged Rules
• Length before Voiceless

/khae pc ph kh ae bc b b ae kc
kh b ae gc g/
14

• Phonological Processes Ladefoged Rules
• voiced ? longer when at end of
  syllable sass, shook vs. push
• stop ? unreleased before stop apt, act
  (often see some mark in spectrogram)
• voiced, alveolar, stop ? glottal stop
  when before an alveolar nasal in same
  word beaten ? /b iy q en/
• nasal ? syllabic at word end when following
  obstruent chasm ? /k ae z em/ NOT film
  (obstruent complete closure of airway /l/ is
  not)
• liquid ? syllabic at word end and following
  consonant paddle, whistle, kennel NOT
  snarl unless classify /r/ as vowel, syllabic

15

• Phonological Processes Ladefoged Rules

/ae pcl tcl th ae
kcl tcl th/
/bcl b iy q tcl
en ax_h/
16

• Phonological Processes Ladefoged Rules
• alveolar, stop ? voiced, flap when
  between two vowels, second of which is
  unstressed This rule has speaker-dependent
  variations
• alveolar, stop ? omitted between two
  consonants most people, sandpaper, grand master
• consonant ? shortened before identical
  consonant
• ? ? voice, stop between nasal and voice,
  fricative when following vowel absent or
  unstressed prince vs. prints (e'penthesis)
• ? ? following word-final nasal,
  consonantal nine come
  sang (e'penthesis)

17

• Phonological Processes Ladefoged Rules
• most people and grand masters use sandpaper

/m ow s pc ph iy pc ph el n gc g r ae n m ae
s tc th er z yu z s ae n pc ph ey pc ph er/
18

• Phonological Processes Ladefoged Rules
• nine come sang

/n ay n ax kcl kh ah m ax
s ae ng ax/
19

• Phonological Processes Ladefoged Rules
• vowel ? longer in open syllables sea vs.
  seed vs. seat sigh vs. side vs.
  sight(equalize length of syllables with
  differing numbers of segments)
• vowel ? longer in stressed syllable below
  vs. billow(stressed syllables are longer in
  duration than unstressed)
• vowel ? nasal before nasal consonant
• vowel, stressed ? schwa (vowel
  reduction) able vs. ability Canada vs.
  Canadian photograph vs. photography

20

• Phonological Processes Ladefoged Rules
• sigh side sight

/s ay s
ay dcl d s a tcl th/
21

• Phonological Processes Ladefoged Rules
• below billow

/b ax l ow
b ih l ow/
22

• Phonological Processes
• Why is this useful? (a) Providing models of
  known phenomenon is better than having
  classifier learn the phenomenon from data
• (b) Provides humans with appropriate cues for
  understanding, naturalness
• (c) Accurate phonetic modeling improves ability
  of classifier to discriminate between classes
• Example for Text-to-Speech (case (b))? Create a
  TTS system? Dont shorten vowels before
  voiceless plosives? Creates, by default,
  acoustic cue for voiced plosives? Decrease
  intelligibility or at least naturalness of system

23

• Phonological Processes
• Example for Automatic Speech Recognition (case
  (c)) ? Train a speech recognizer using
  dictionary pronunciation ? Then, in all cases
  where voice, stop between nasal and
  voice, fricative such as fancy (in CMU
  dictionary as /f ae n s iy/), acoustics show
  alveolar stop, but trained as either nasal /n/
  or fricative /s/. ? Decreases ability of
  model to discriminate classes ? Decreases
  performance of system
• Difficulty is in providing comprehensive,
  accurate rulesthat are not inappropriately
  forced on a system

24
Stops/Plosives

• There are six plosives (oral stops) in American
  English
• . bilabial alveolar velar
• unvoiced /p/ /t/ /k/
• voiced /b/ /d/ /g/
• plus the flap /dx/ which is a very short /t/
  or /d/
• Plosives can be difficult to identify and
  discriminate contextual cues can be varied
• Cue (1) is the formant transitions of
  neighboring vowels
• for bilabials, F2 drops at CV boundary
• for alveolars, F2 goes toward 1800 Hz at CV
  boundary
• for velars, F2 may meet F3 (velar pinch) or be
  fairly flat
• Cue (2) is that voiced plosives may have
  pre-voicing more likely when plosive is
  between two vowels

25
Stops/Plosives

• Cue (3) is that voiced plosives usually have VOT
  of lt 30 msec,
• but unvoiced plosives usually have VOT of gt 50
  msec
• Cue (4) is that the VOT is shortest for
  bilabials, longer for alveolars, and longest for
  velars. (VOT /p/ lt /t/ lt /k/ and /b/ lt /d/ lt
  /g/)
• Cue (5) is that aspirated (unvoiced) plosives
  show evidence of F2 and F3 during aspiration
  voiced plosives usually dont
• Cue (6) is the spectral shape in theory, the
  shape of the spectrum at burst release can be
  used to distinguish plosives
• /p/ and /b/ have energy low in frequency or
  weakly spread throughout spectrum,
• /t/ and /d/ have more energy above 4KHz (related
  to alveolar fricatives /s/ and /z/),
• /k/ and /g/ tend to have more well-defined peaks
  in the spectrum (near formant locations).

26
Stops/Plosives

• Other cues related to spectral shape
• Cue (7a) In the context of front vowels, /k/
  and /g/ have
• spectral peak just above F2 of adjacent vowel,
  making them
• confusable with /t/ and /d/ but front vowels
  show more
• velar pinch
• Cue (7b) In the context of back vowels, /k/ and
  /g/ have one
• spectral peak between 1000 and 1500 Hz, a second
  peak
• between 3000 and 4500 Hz.
• Cue (8) Velar bursts also sometimes display
  double burst,
• or a second burst during the frication
• Cue (9) Post-vocalic consonants are often
  unreleased they can be identified by (a)
  glottalization, (b) sudden drop in vowel
  energy, or (c) formant movement at end of vowel

27
Stops/Plosives

• Cue (10) When the plosive is unreleased, the
  voicing distinction is based more on length
  of preceding vowel voiced plosives are
  associated with longer vowels, unvoiced plosives
  with shorter vowels
• Cue (11) In V1C1C2V2 patterns, where both C are
  plosives, the existence of two plosives is in
  the different formant transitions in V1 and
  V2, the longer duration of closure, and sometimes
  in a brief
• click in spectrum indicating a change in
  place of articulation
• Cue (12) Plosives have different
  characteristics in stressed vs. unstressed
  environments. VOT for unvoiced plosives before
  unstressed vowels is shorter than VOT for
  unvoiced plosives before stressed vowels
  plosives in an unstressed-vowel environment
  are less spectrally clear in unstressed
  syllables, /t/ and /d/ may be realized as a
  flap /dx/.

28
Stops/Plosives

• Cue (13) Flaps have short duration (lt 30 msec),
  dip in energy levels between two vowels, weak
  F2 and F3, and F2 tends toward
• 1800 Hz
• Cue (14) Consonant clusters can provide
  restrictions for 3-consonant clusters
  (beginning with /s/-plosive), the only valid
  combinations are /s p l/, /s p r/, /s p y/ /s
  t r/, /s t y/ /s k l/, /s k r/, /s k y/, /s k w/
• Cue (15) In /s/-plosive-vowel combinations, VOT
  tends to be shorter and duration of /s/
  shorter than normal

29
Plosives Unvoiced Initial in Front-Vowel Context
/p iy t iy
k iy/
30
Plosives Voiced Initial in Front-Vowel Context
/b iy d iy
g iy/
31
Plosives Unvoiced Initial in Mid-Vowel Context
/p ah t ah
k ah/
32
Plosives Voiced Initial in Mid-Vowel Context
/b ah d ah
g ah/
33
Plosives Unvoiced Initial in Back-Vowel Context
/p aa t aa
k aa/
34
Plosives Voiced Initial in Back-Vowel Context
/b aa d aa g
aa/
35
Plosives Unvoiced Final in Front-Vowel Context
/iy p iy t
iy k/
36
Plosives Voiced Final in Front-Vowel Context
/iy b iy d
iy g/
37
Plosives Unvoiced Final in Mid-Vowel Context
/ah p ah t
ah k/
38
Plosives Voiced Final in Mid-Vowel Context
/ah b ah d
ah g/
39
Plosives Unvoiced Final in Back-Vowel Context
/aa p aa t
aa k/
40
Plosives Voiced Final in Back-Vowel Context
/aa b aa d
aa g/