Fundamental Frequency and the Production of Intonation in SSBE

1
Fundamental Frequency and the Production of
Intonation in SSBE

• Meg Zellers
• Research Centre for English and Applied
  Linguistics, University of Cambridge
• Sound to Sense Project 8 Prosodic Structure and
  Fine Phonetic Detail

2
Why bother with prosody?
3
Prosody? Intonation?

• Prosody sound phenomena on a level higher than
  the segment
• Lexical stress
• Pauses
• Voice quality
• Intonation
• Intonation subset of prosody dealing with pitch
  change
• Tone languages pitch is part of lexical identity
• Non-tone languages pitch has other purposes

4
Defining Pitch

• A perceptual construct to do with the way we
  perceive fundamental frequency (F0)
• In speech, the overall frequency of a complex
  wave

5
Problems with Pitch

• Linear change in F0 ? linear change in perceived
  pitch

200 Hz to 400 Hz
400 Hz to 600 Hz
6
Problems with Pitch

• By definition, F0 requires a periodic sound, but
  not all speech sounds are periodic

7
Problems with Pitch

• Other factors besides F0 itself can contribute to
  our perception of pitch
• Intrinsic pitch of vowels a high vowel (e.g.
  i) will sound higher than a low vowel (e.g.
  a) at the same F0
• Voice quality creaky voice can make a sound at
  the same F0 sound lower
• So must account for differences between
  measurable signal and what listeners perceive!

8
How do we use pitch?

• Biological codes (cf. Gussenhoven 2004)
• Frequency code (Ohala 1983) high pitch
  correlates with small sizeExample
  authoritativeness vs. submissiveness
• Production code high pitch correlates with more
  available energyExample pitch reset at
  sentence beginnings
• Effort code larger pitch movements correlate
  with more exertionExample larger pitch
  excursions when surprised
• Sociolinguistic uses variation in choice of
  pitch range to create gender identity

9
How do we use pitch?

• Some aspects of intonation can be described in
  the form of a grammar 2 groups of theories about
  structure
• Contour-based approaches
• Traditional British approach (Cruttenden 1997)
• IPO approach for Dutch (t Hart et al. 1990)
• Target/tone-based approaches
• Autosegmental-Metrical theory (e.g. Pierrehumbert
  1980)

I have a present for you.
10
Intonational Structure

• Intonation applies over the domain of the
  Intonational Phrase (IP), defined internally by
  presence of at least one pitch accent (nuclear
  accent)
• Pitch accents placed on prominent syllables
• Boundary tones at phrase edges

I have a present for you. L HL
L
11
Intonational Structure

• Finite-state grammar of intonation (Pierrehumbert
  1980)

12
Intonational Structure

• IP boundaries normally fall at syntactic
  boundaries
• BUT the presence of an appropriate syntactic
  boundary doesnt guarantee the presence of an IP
  boundary
• Pauses may occur at IP boundaries
• BUT do not necessarily have to (and may occur at
  non-boundaries e.g. in speech errors)
• If no pause, other signals still present

13
Contribution to language

• Higher level than lexical identity
• Segmentation and parsing
• Babies can use intonation from an early age,
  possibly to aid in segmentation for acquisition
  (Fernald 1984)
• Adult listeners use intonation to parse
  (garden-path effect, Blodgett 2004)
• But intonation not strictly necessary for
  segmentation/ parsing!
• Uses higher than the sentence level discourse
  construction

14
Discourse Structure

• Discourses are units because they are made up of
  parts which fit together coherently
• Listener assumes that message to be interpreted
  will be coherent, and looks for clues as to how
  pieces fit together (Brown Yule 1983)

He was standing in the store when a woman came
in. He was standing in the store when the woman
came in.
15
Focus and Intonation

• Focus is one aspect of discourse structure that
  has been widely studied w.r.t. intonation
• An ambiguous concept semantically, selection of
  one item from a group of competitors (Steedman
  2003) could also be a prosodic tool used to mark
  a variety of semantic phenomena (J. Local,
  personal communication Selkirk 1995)
• Focus can mark distinction between given and new
  information in an utterance

16
Focus and Intonation

• In e.g. Neapolitan Italian, timing of F0 peaks
  varies systematically in relationship to focus
  focused elements in questions have later F0 peaks
  (cf. DImperio 2000)

17
Topic Structure

• Topic structure is one kind of discourse
  structure occurring on a level higher than single
  sentences
• Topic is a fuzzy concept what the unit of
  discourse is about. Could be an entity, a
  proposition, etc.
• Hard to know what to call a topic structure unit
• Intonation sometimes uses paratone, by analogy
  from paragraph
• But connotations can be problematic!
• Topic unit is a more neutral term for now

18
Topic Structure

• Grosz Sidner (1986)
• Topic structure simple hierarchy
• Discourse Segments have Discourse Segment
  Purposes (DSPs)
• DSPs contribute to overall Discourse Purpose
• Topic would have to be mapped to one level in
  an arbitrary way could be weakness or strength
  of this theory

19
Topic Structure

• Wichmann (2000)
• 4 levels of topic structure, defined by semantic
  relationships Topic, Continuation,
  Elaboration, Reformulation

• Nakajima and Allen (1993)
• Similar to Wichmann, but with slightly different
  categories Topic Shift,Topic Continuation,Elabo
  ration,Speech Act Continuation

• Main difference is in order of ranking of
  categories

20
Prosody and Organization

• No 11 correlation between prosodic marking and
  topic structure organization
• Discourse structure is ambiguous, which can
  confuse pattern observation
• Despite this, clear patterns
• Focus and peak timing
• Deaccentuation of given information (Baumann
  Grice 2006)
• Pause length correlation with boundary strength
  (Swerts 1997)
• F0 peak height and timing correlation with topic
  structure (Wichmann 2000)

21
Experiment Motivation

• Intonation research is trying to discover
• abstract structure of intonation
• how physical signal maps to that structure
• Wichmanns theory about topic structure
  categories and intonational cues provides way to
  test this
• Production Code versus linguistic system

22
Experiment Methodology

• Written text with items falling into four topic
  structure categories
• Topic (new topic)
• Addition (more info, same topic)
• Elaboration (more detail or clarification of
  preceding utterance)
• Continuation (completion of a speech act started
  in the previous utterance)

23
Experiment Methodology

• Topic structure categories interact with other
  factors that could influence intonation
• Position in sentence (beginning or end)
• Position within utterance group (declination over
  time)
• Presence/absence of anacrusis (unstressed
  syllables preceding stressed target vowel)
• Identity of stressed vowel

24
Experiment Methodology
25
Experiment Methodology

• 3 sets of target words (3 syllables with initial
  stress)
• Stressed vowels /e/, /æ/, and /I/ respectively
• Up to 90 possible tokens per speaker

University was always a desire for Emory. He was
sure that without an education, minimal
opportunities would be in his reach. To this
end, he studied constantly. At a minimum he read
a book a week.
26
Experiment Methodology

• 20 subjects, native speakers of Standard Southern
  British English (SSBE)
• Recordings annotated for intonation contour (only
  HL used in analysis) and other F0
  characteristics
• Items in Final position in sentences did not vary
  systematically, so only results for Initial and
  Anacrusis sentence position are shown here

27
Experiment Hypotheses

• In all cases, that intonational factors would
  vary on basis of topic structure, with TgtAgtEgtC
• F0 peak timing peaks more delayed going up
  hierarchy (so T peaks timed latest) (cf. Wichmann
  2000)
• F0 peak height peak height above speaker
  baseline increases going up hierarchy (cf.
  Wichmann)
• F0 fall range pitch range of F0 falls increases
  going up hierarchy
• F0 fall slope slope of F0 falls increases going
  up hierarchy

28
Experiment Results

• F0 peak timing
• Failed to replicate results of Wichmann (2000)
• T only timed later in Initial condition
• Results for other categories ambiguous
• May be too many confounding factors due to other
  uses?

29
Experiment Results

• F0 peak height measured as height in semitones
  (st) from speakers baseline pitch
• Significant correlation with position in
  utterance group (1gt24gt5), but not to topic
  structure

30
Experiment Results

• F0 fall range distance between H and following
  L valley in semitones
• Significant correlation with topic structure
  (TgtAEgtC), but possible interaction with position
  in group

31
Experiment Results

• F0 fall slope range divided by time between H
  and L
• Significant correlation with topic structure
  (TgtAEC), but again possible interaction with
  position in group

32
Experiment Discussion

• Range and slope data support the hypothesis that
  speakers vary F0 to signal discourse structure
• New topics privileged
• Intermediate level for Additions/Elaborations
• Data ambiguous as to status of 2 different
  theories of topic structure could fit either
• Hierarchy T is always at the top of a hierarchy,
  and C is always embedded under another unit A
  and E simply fall in the middle
• Categories E and A are one category, not two

33
Experiment Discussion

• Range and slope may be arbitrarily structured and
  specified in language system however, peak
  height seems to be based on universal principles
  (Production Code).
• Might have been forced due to way in which text
  was presented to readers
• A tricky boundary to navigate due to overlap in
  positional and (possibly) categorical
  considerations

34
Future Directions

• Perception experiment to what extent are
  listeners attuned to these signals?
• Cross-linguistic investigation are these cues
  used universally, or are they specified by the
  languages grammar?
• Peak timing very sensitive in other conditions
  (e.g. focus) perhaps too many confounds here

35
Conclusion

• Despite being a half-tamed savage (Bolinger
  1978), intonation does have many linguistically
  structured aspects
• F0 modulation, while not the only tool available,
  is useful in a variety of ways for speakers as
  they shape their utterances into a coherent
  discourse
• It remains to be seen how this interacts with
  other prosodic characteristics to create the rich
  expressive system that is available to us in
  spoken language

Thank you!