Neural functional organization account of speech perception

1
Distributed auditory processing is compatible
with the information conveyed by the speech signal
Sarah Hawkins and Ingrid Johnsrude Phonetics
Laboratory, Department of Linguistics, University
of Cambridge
sh110_at_cam.ac.uk ingrid.johnsrude_at_queensu.ca
Supported by the Leverhulme Trust, CIHR,NSERC
and Canada Research Chair Progamme
is consistent with a distributed, interactive

• Problem Models of speech perceptionoften
  emphasize phonetic or phonological
  categories(features, phonemes, gestures) that
• are stable, abstract entities
• result from stripping irrelevant variation from
  speech
• are prerequisite to the processing of other
  aspects of speech (grammar and meaning).

• Neural functional organization account of speech
  perception
• Distributed
• Speech comprehension probably relies on multiple
  cortical networks that operate in parallel
• This functional organization may map onto
  anatomically segregated processing streams
  similar to those identifiedin macaque monkeys.21
• Functional neuroimaging results are consistent
  with multiple, parallel, cascaded auditory
  streams of processing.22,23,26-28
• The subcortical auditory system is also highly
  parallel.
• Neurophysiological studies suggest that
  information ineven core auditory cortex regions
  is integrated over many time domains
• This suggests that multiple representations of
  the input,at many temporal grains, are
  simultaneously available for processing by higher
  centres.32,33
• Interactive
• Information flow in the auditory system is not
  unidirectional. Eachcortical feedforward
  connection has its feedback complement.29-31
• Anatomy suggests converging influences from
  multiple higher stages of perception on lower
  stages.21
• Auditory cortical responsivity is context
  dependent and plastic, and
• probably driven by feedback from higher-order
  areas.
• Behaviourally relevant stimuli produce rapid
  changes in the response characteristics of single
  units in ferret primary auditory cortex.39
• Humans demonstrate greater fMRI activity to
  learned phonological contrasts in auditory belt
  or parabelt.40

Examples Grammatical information conveyed by
systematic acoustic variation b. About
Morphemes

• a. About function vs content words Pronounbe
  system
• Speakers show different assimilation in function
  and content words. E.g. /m/ assimilates to place
  of next consonant in Im but not lime or crime11
• Im blowing / going / watching a?m
  a?? a?w lime bark lime goes crime wave
  a?m
• The acoustic pattern may be used by the listener
  to inform about the grammatical class of the
  speech segment being perceived.In its place in
  an utterance, Im has few or no acoustic
  competitors.

• Concluding remarks
• Fine phonetic detail simultaneously informs about
  perceptual units at multiple linguistic levels
  and thus over different time domains.12,13,35
  Compatible with multiple, parallel,
  hierarchically organized anatomical pathways.
• Each linguistic category is relational and
  plastic each is bound with other elements
  (larger, smaller) and no element can be described
  independently of its prosodic, grammatical, and
  functional context. Supported by interactive
  organization of stages of auditory processing.
• Rapid perceptual tuning is manifest in many ways,
  e.g., flexible phonemic category boundaries and
  perceptual learning of degraded speech.1-9 May
  rely on feedback from higher areas to early
  auditory regions.
• Models of visual object perception posit Bayesian
  integration of image features and knowledge to
  determine the most probable interpretation of the
  current input. eg 38,41,42,43,44 A similar
  mechanism may act to combine information from
  multiple sources (top down and bottom up) to
  constrain speech perception.
• The framework outlined here
• requires that no unit of speech perception is
  primary
• emphasizes the knowledge-driven nature of
  perception
• coherently integrates behavioural,
  neurobiological and linguistic data and theory

BUT this over-emphasizes phonology
early loss of detail seems unlikely

• Phonemic category boundaries are context
  dependent thus plastic
• Boundaries shift with phonetic context, stimulus
  distribution and range, meaning (lexical status,
  sensible world) shifts are fast1-9
• Much phonetic variability is systematic and
  informs about properties other than phonemic
  categories
• Realization of phonemes is systematically
  influenced by10,11
• 1) allophonic variation
• position in the syllable (eg tip vs pit)
• boundaries between words (eg grey train vs
  great rain)
• grammatical status (eg productivity of a
  morpheme content vs function words)
• 2) speaker intent register (discourse
  function, casualness, rate)
• 3) talker identity
• Experiments show listeners use much of this
  variability12-17

Some of the perceptual information available from
/m?st/ in mistimes and mistakes
Perceptual information available in the short
sound section mist, fromTess mistimes it and
Tess mistakes it. Information about featural,
phonemic and lexical identity, and syllabic,
morphemic and grammatical structure, is conveyed
simultaneously in the fine phonetic detail
events at segment boundaries and longer-term
relationships. Prior knowledge is required for
linguistic informationat all levelsto be
extracted from sensory input. No unit is
identifiable independent of context, and no
unit/level is primary. Information is mapped onto
prosodic structures linked to grammatical
structures12,13,34 . Examples of structures are
available at http//kiri.ling.cam.ac.uk/sarah/doc
s/CNS06trees.pdf
Bold font nodes in linguistic structure
potential perceptual units
For references, see separate handout or email us