Why an objective intelligibility assessment ?

1
Why an objective intelligibility assessment ?
Catherine Middag Jean-Pierre Martens
Gwen Van Nuffelen Marc De Bodt
2
Speech intelligibility

• The degree to which a listener understands the
  auditory signal produced by a speaker
  
• Duffy 2007
• Different linguistic levels phoneme word
  sentence connected speech
• Clinical relevance?
• Model of Functioning, Disability and Health
  (WHO, 2001)

3
health condition e.g. dysarthria
body functions and structures phonatory,
articulatory, resonatory and respiratory muscles
activity speech intelligibility
participation
Index of severity
environmental factors
personal factors
4
Clinical relevance

• depends upon the main subsystems of speech
• articulation
• resonance
• phonation
• prosody / respiration
• 2. an index of severity of functional limitations
• to document the disorder
• therapy outcome measure
• monitoring change in performance over time
• 3. gives an idea of adequacy of communication
  the impact on participation in daily life

5
Golden standard

• Perceptual intelligibility assessments
• Scaling rough overall index
• Measurements () more accurate index, procedures
  and test materials can be standardized

Important for reliability and validity
6
Reliability and validity
speaker message
transmission system
listener
influencing factors
target
7
Reliability and validity

• transmission system easy to control /
  standardize
• listener intelligibility score depends upon the
  listeners
• familiarity with speaker/ speakers accent/
  cultural background/ pathology
• familiarity with test items
• test items predictability
• solutions

8
Reliability and validity
large pool of test items random selection
e.g. Assessment of Intelligibility of Dysarthric
Speech (Yorkston, 1981) non-existing words
syntactically correct sentences conveying no
meaningful message e.g. Swedish
Intelligibility Test (Lillvik, 1999) Artificial
Negative effect on naturalness Dutch adult
patients Dutch Intelligibility Assessment (De
Bodt et al. 2006)
9
DIA

• Intelligibility at phoneme level
• 50 consonant vowel consonant words
• 3 subtests
• A initial consonants (19 words)
• B final consonants (15 words)
• C medial vowels/ diphthongs (16 words)
• Measures to improve reliability
• Balanced mix of existing and non-existing (well
  pronounceable) words
• Large pool of test items 25 lists of words /
  subtest
• Random selection
• In each list one randomly selected phoneme
  occurs twice

10
DIA
List A10
11
Reliability and validity

• Advantage of a speech technology based
  intelligibility assessment
• Computer replaces human listener
• No need for a large pool of test items and random
  selection
• No need for artificial features needed

12
Space intelligibility tool

• acoustic models of phonemes ? phonemic features
• and/or
• acoustic models of phonemes ? phonological
  features

articulatory dimensions e.g. voice, place of
articulation, manner of articulation, if these
features are able to predict intelligibility good
chance they can characterize the articulation of
a pathological speaker
13

• Thank you!