Invariance

LING 285
Spring 2021
Mary Byram Washburn
 Speech Speech
Synthesis Recognition

Speech
Recognition

Challenge: Challenge:
Coarticulation Invariance
 Which of these is the word “ship”?
1 2 3 4 5

ship

Invariance
 Which of these is the word “ship”?
1 2 3 4 5

ship

Invariance
 Which is the word “ship”?

ship

Invariance
 The Problem of Invariance
 The acoustic features of a word are never the same from one
utterance to another
Invariance
 Interspeaker Variation
 Each speaker has unique acoustic features

Car?
Bag?
Interspeaker  What is your generic term for a sweetened carbonated beverage?
Variation
Interspeaker
Variation
Interspeaker  pecan
Variation
Interspeaker
Variation
  A Study of the Social Dialects in Detroit
 Roger W. Shuy (1968)

Interspeaker
Variation
  Interview
 700 informants across Detroit in 1966
 1.5 hrs

 Interested in who used a variety of different features.

Interspeaker  Full Nasal vs. Deletion of Syllable Final Nasal
Variation:  If deleted, replaced with a nasalized vowel.

Shuy 1968  an’ he started vs. ã he started

 come down vs. cõe dõw̃
 man from Uncle vs. mã frõ Ũcle
 Social Class Sex
20 20
Percent of Nasals Deleted

Percent of Nasals Deleted

16 16
12.9 12
12 12 11.1
9.5
8 6.8 8
5.5
4 4
0 0

I II III IV Female Male

Age Race
Percent of Nasals Deleted

Percent of Nasals Deleted

20 20

15 15 14.2
12.3
9.7 8.7
10 10
6.6
5 5

0 0

Adult Teenager Child Black White

 Social Class and Race
20 19.3
18
14.8
Percent of Nasals Deleted
16
14 13.3
12 11.3
10
8 6.3
6 5.3 5
4
2 0
0
Black White
I II III IV
 Social Class and Age
20
Percentage of Nasals Deleted 18
16
13.9
14 12.8 12.7 Adult
12 11.4
Teenager
10
Child
8
6 5.4 4.6
4
2
0
Middle Class Working Class
 The Problem of Invariance
 The acoustic features of a word are never the same from one
utterance to another
Invariance
Geographic Region
 Interspeaker Variation Age
 Each speaker has unique acoustic features Social Class
Race
Gender
 Who is the
Interspeaker system built to
understand?
Variation
 The Problem of Invariance
 The acoustic features of a word are never the same from one
utterance to another
Invariance
Geographic Region
 Interspeaker Variation Age
 Each speaker has unique acoustic features Social Class
Race
What, specifically, is different?
Gender
 Acoustic differences in the word “car”?

Adult Female Boston Northern

Male Child Cities
  Phoneme:
 Minimal unit of sound

Phonemic
 Phonemic Inventory
Inventory  Set of phonemes a speaker uses
 Different for different speakers
 ex: Different for different languages
Phonemic  Phonetic Inventory: sounds of a language
 Every language has a unique phonetic inventory.
Inventories
 How do you determine whether a
phoneme exists in your speech?
Minimal Pairs

 Phoneme:
 Minimal unit of sound

Phonemic
 Phonemic Inventory
Inventory  Set of phonemes a speaker uses
 Different for different speakers
 ex: Different for different languages, different dialects

Eastern American: /ɔ/ Southeastern American: no /ɛ/ Detroit, Working Class,

Dawn: /dɔn/ Pen: /pɪn/ Black, Male, Teenager
Don: /dɑn/ Pin: /pɪn/ down: /dɑ̃ ʊ̃/
  Minimal Pairs
Phonemic  Words that differ by only one phoneme
Inventory
  Minimal Pairs?
 capevs. Kate
 ridervs. writer
 to bow vs. a bow
 lei vs. ray
 baked vs. bagged
Minimal Pairs
 Write Minimal Pairs for these phonemes?
 /p/ vs. /b/
 /s/ vs. /z/
 /n/ vs. /ŋ/
 /θ/ vs. /ð/
  Minimal Pairs
 Words that differ by only one phoneme
Phonemic
Inventory  If changing the phoneme changes the word
 phonemes are contrastive (different)!
  Actual acoustics of a phoneme varies…

 pa /pɑ/
 pop /pɑ/
Phonemes
 /phoneme/
[actual acoustics]
 Actual acoustics of a phoneme varies…

 pa /pɑ/ [pɑ]
 pop /pɑ/[pɐ̗]
Phonemes
  Actual acoustics of a phoneme varies…

 pa /pɑ/ [pɑ]
Phonemes  pop /pɑ/[pɐ̗]

 top /tɑp/
 stop /stɑp/
  Actual acoustics of a phoneme varies.
 Accept a range of acoustic signals for a phoneme

Phonemes

t
  Actual acoustics of a phoneme varies…

 pa /pɑ/ [pɑ]
Phonemes  pop /pɑ/[pɐ̗]

 top /tɑp/ [thɑp]

 stop /stɑp/ [stɑp]
aspiration
  Aspiration = big difference acoustically!
 tock /tɑk/ [thɑk]
 stock /stɑk/ [stɑk]

As big as tock vs. dock

Phonemes

tock stock dock

  Actual acoustics of a phoneme varies.
Contrastive
Phonemic

Phonemes Noncontrastive

/t/ /d/
  Actual acoustics of a phoneme varies.
 Accept a range of acoustic signals for a phoneme
 What that range is depends on your language/dialect.

 English
 [thɑl] tall
 [tɑl] Aspiration isn’t contrastive.
Minimal Pairs (same phoneme with or without aspiration)
 [sthɑl] stall
 [stɑl]

 Korean
 [tal] 달 moon Aspiration is contrastive.
 [thal]탈 mask (different phoneme!)
 [thɑk] [stɑk] [dɑk]
Aspiration is
noncontrastive.

Voicing is
phonemic.

Phonemes
Aspiration is
contrastive/
phonemic.

[tal] [thal] [t*al]

달 탈 Tensed
moon mask 딸
daughter
 tuɛidifɑiviz
pronunciation English pronunciation English
translation translation

1 [phif] horse [pif] look

2 [nɛm] house [nɛn] house

Minimal Pairs 3 [fop] door [pof] cup

4 [fin] tissue [fɛn] spoon
5 [sop] snow [top] printer

Phonemes?
  Actual acoustics of a phoneme varies.
 What is contrastive depends on your language/dialect.

Contrastive
Phonemic

Phonemes Noncontrastive

/t/ /d/
 New All 4 are
England different 4 phonemes
words.

1. /dɔn/
2. /dɑn/ 1&2 are
different
Phonemic Southeast
US
words.
3 phonemes
Inventory 3. /pɛn/ 3&4 are
the same
4. /pɪn/ word.

1&2 are
the same
West Coast word.
US 3&4 are
3 phonemes
different
words.
 The Problem of Invariance
 The acoustic features of a word are never the same from one
utterance to another

 Interspeaker Variation Dialectology

 Each speaker has unique acoustic features and unique sets of
phonemes for a given word
Invariance  Speakers have different phonemic inventories
 Each speaker accepts a different range of acoustic values for a phoneme
 Phonemic Inventory:
 Milwaukee, WI Front Vowels

/ɪ, e, a/

 Los Angeles, CA
/i, ɪ, ɛ, e, æ/
Dialectology

 Conway, AR

/i, ɪ, e, æ/
 Phonemic Inventory:
 Milwaukee, WI Consonants

White Black
/t, d, θ, ð/ /t, d/
/læst/  /læs/

 Los Angeles, CA

Dialectology White Black, Hispanic

/t, d, θ, ð/ /t, d/
/læst/  /læs/

 Conway, AR

Higher Class Lower Class

/t, d, θ, ð/ /t, d/
/læst/  /læs/ /læst/  /læs/
 The Problem of Invariance
 The acoustic features of a word are never the same from one
utterance to another

 Interspeaker Variation
Dialectology
 Each speaker has unique acoustic features and unique sets of
Invariance phonemes for a given word
 Speakers have different phonemic inventories
 Each speaker accepts a different range of acoustic values for a phoneme

 Intraspeaker Variation Phonology

 An individual speaker pronounces a phoneme different depending
on the phoneme’s environment

How do humans do it…?

 In what
peek environment do keep
[phik] American English [khip]
speakers use a lot
of aspiration?

Phonology: Phonological Rule

Aspiration
Rules
 Word IPA Environment
peek [phik] Word initial, before [i]
NOT word final, after [i]
keep [khip]

totes [thots]

Phonology:
Aspiration
Rules
Phonological Rule for Big Aspiration:
 Word IPA Environment
peek [phik] Word initial, before [i]
NOT word final, after [i]
keep [khip]

totes [thots]

Phonology: peeking [phi.khiŋ]

Aspiration
Rules
Phonological Rule for Big Aspiration:
 Speech Recognition:
- Range of acceptable acoustics to represent a phoneme
- That range will be different for every speaker

The Problem of Invariance

 The acoustic features of a word are never the same from one
utterance to another

 Interspeaker Variation
Dialectology
 Each speaker has unique acoustic features and unique sets of
Invariance phonemes for a given word
 Speakers have different phonemic inventories
 Each speaker accepts a different range of acoustic values for a phoneme

 Intraspeaker Variation Phonology

 An individual speaker pronounces a phoneme different depending
on the phoneme’s environment
Jobs for
Linguists Panel