Suprathreshold Speech Recognition I Flashcards Preview

HESP606 > Suprathreshold Speech Recognition I > Flashcards

Flashcards in Suprathreshold Speech Recognition I Deck (22):
1

Describe the purposes of suprathreshold speech recognition testing. What does each assume? What do both purposes incorporate?

  • Assess precision of speech transmission system (metro)
    • Purpose at Bell Labs, Air Force
    • Assumes an ideal observer
      • assumes young normal, paying attention
  • Assess precision of individual’s speech reception
    • Assumes an ideal transmission system
    • Clinical implications:
      • Diagnostic purposes (determining site of lesion)
      • Estimating functional deficits of hearing loss
  • both purposes incorporate understanding of speech
    • Using suprathreshold presentation to determine how well an individual performs

2

Describe the requirements of a speech recognition test. What should the material be representative of? Why? What else should the test be? How difficult should the materials be? How difficult are different materials?

  • Material should be representative of the language
    • Concept of validity: performance should reflect the person’s ability to understand every day speech
    • Use material that represents the language (face validity)
    • Test materials developed to reflect the sounds in average speech - use phonemes in the language
  • test results should be reliable (repeatable)
  • test should be easy to score
    • Connected discourse: difficult to score
    • Word lists: easy to score
  • materials should be relatively difficult to obtain a measure of the person’s speech intelligibility problem
    • Nonsense syllables –too abstract, too difficult for some people, measure absolute recognizability rather than intelligibility  
    • Monosyllabic words –not abstract, not too much redundancy; a measure of intelligibility
      • Need to hear all phonemes to get it right
    • Multisyllabic words –scores are influenced by knowledge of the language (lexicon)
      • Stress patterns, hearing syllables only
    • Sentences –knowledge of syntax, semantics, and linguistics affects performance

3

Describe the history of monosyllabic word tests. What was the first attempt to develop a speech test called? What was its purpose? What were the four requirements?

  • Harvard Psychoacoustics Labs(Harvard PAL, Egan et al., 1940’s)
    • 1st attempt to develop a test of speech recognition
      • Called Harvard PAL PB-50’s
    • purpose for test development: assess communication
      • System or individual
    • first requirement: wanted a valid test
      • phonetic balancing scheme –individual phonemes occur as frequently on the list as they do in the language
      • Phonetic balancing for initial phonemes only
      • Based on published data of phonetic composition of 100,000 words
    • second requirement: monosyllabic in structure
      • measure individual’s ability to discriminate among fine elements of speech (multisyll. words –too easy; nonsense syll–too difficult)
    • third requirement: lists of equal difficulty
      • Concept of list equivalence
      • Range of difficulty on each list
      • Need to produce similar scores across lists 
    • fourth requirement: words should be in common usage
       

4

Describe the structure of the PAL PB50 test. How many lists and words per list? What were later recordings of the test? How is it most widely used?

  • Harvard PAL PB50’s (continued)
    • 20 lists, 50 words/list
    • never recorded at Harvard
    • later recording:
      • Rush Hughes was talker
      • Very difficult test
    • most widely used as MLV
      • carrier phrase “say the word …”; rationale
        • To have a stable sample of language to emned test word in
      • Peak a word in the carrier phrase at 0 VU
        • Not test words
      • scoring: write-down, talk-back, on-line recording
        • Hard to tell what they say
           

5

Describe the CID W-22 PB lists. What were the problems of the PAL PB-50's? What were the criteria? How many lists were developed? Are the lists equivalent?

  • CID W-22 PB’s (Hirsh, 1952)
    • problems with PAL PB-50’s
      • Poor recording, unfamiliar words, imperfect phonetic balance
    • criteria:
      • monosyllables
      • Highly familiar words
      • phonetic balance  -based on printed and spoken English (Initial, Medial, Final positions)
        • Spoken - from telephone recordings from Bel labs
    • developed 4 lists, 50 words/ea
      • recordings with Hirsh as speaker
      • List 1 easier than Lists 2, 3, 4
      • high test-retest reliability

6

Describe psychometric functions. Where are the linear and curved segments? What should be analyzed? What does a steeper slope represent? What are the problems with CID W-22?

  • 50% at 14 -> Linear from 10 to 20 dB HL.
  • Analyze slope at linear segment - 5% per dB 
  • Analyze level of Max performance
  • Steep slope is easier**
  • CID W-22 is not equivalent and too easy. 

A image thumb
7

Describe the NU-6 Test. What was the goal of the test development? What scheme is used? What were the words structured like? How many lists did they develop? What kind of speakers did they use in the recordings?

  • Northwestern University Test No. 6
    • Tillman & Carhart, 1966
    • goal for test development: list equivalence
    • scheme of phonemic balancing - perception of phonemes
      • Based on frequency of occurrence of phonemes within their set of words, not on English
    • corpus of words (CNC; Lehiste & Peterson list)
      • consonant nucleus consonant
      • Based on how often occurred on Peterson list (1263 words)
    • developed 4 lists; 50 words/list
    • recordings made (male & female talkers)
      • Popular recording was with Tillman as talker

8

Describe the psychometric functions of the NU-6. For normal hearing and conductive hearing loss. For people with sensorineural hearing loss. Are the lists equivalent?

  • Psychometric functions
    • Normal hearing and conductive hearing losses:
      • slope: 6%/dB; 100% score at 24 dB SL re: SRT
        • normal - 24 dB HL, 44 dB SPL
    • people with sensorineural hearing loss:
      • slope: 3.5%/dB; plateau at 40 dB SL re: SRT
      • Mean plateau was 92% correct
  • Good inter-list equivalence
     

9

How do NU-6 results compare with a female talker?

  • New recordings on CD made by VA with female talker
  • Evaluated by Wilson, Zizz, Shanks & Causey (1990)
    • Compared performance of normal hearing listeners with this recording and with original Tillman recording in quiet and in noise
    • Female talker: slope (Q) = 4.5%/dB (more gradual) ; plateau at 56 dB SPL (36 dB HL), but plateau was not 100% - harder
    • Compared to male talker: slope is similar, but function for female talker is shifted 11-15 dB higher than for male talker
    • Thus, different psychometric functions obtained with different talkers

10

Describe the effects of MLV vs. recorded materials on performance of PB wordlists. What are the advantages of recorded materials? What are the advantages of MLV? What are the disadvantages of MLV? What does ASHA recommend? What does research show?

  • MLV vs. recorded materials
    • advantages of recorded materials
      • Easier to administer (words peaking correctly)
      • Standardized –can compare results between clinics
      • More reliable
    • advantages of MLV
      • More flexible presentation
      • faster
    • disadvantages of MLV –talker differences
    • ASHA recommendations –recorded preferred
    • research findings -inconsistent
       

11

Describe the effects of presentation level on performance of PB wordlists. What is the goal? What is the only way to know you are measuring PB-max? How do you compromise? What are procedures that are not recommended for presentation level? Whare procedures that are recommended for presentation level?

  • Presentation Level
    • Goal: often want to measure best performance
    • Performance-intensity function (PIF) –only way to be sure that you’re measuring PB-max
      • Limitations –takes time
      • Compromise –estimate level needed for PB-max
    • Procedures that are not recommended:
      • MCL ≠ level needed for PB-max
        • underestimates by 20 dB
      • fixed presentation (sensation) level above SRT –
    • Procedures that are recommended
      • fixed SPL (95 dB SPL)
      • just below LDL
      • adaptive procedure to estimate level
         

12

Describe Kamm, Morgan, and Dirks's study on word recognition. What did they compare? Who were their subjects? What were their stimuli? What were their methods? What were their results?

  • Kamm, Morgan & Dirks, 1983
    • Compared word recognition performance obtained with two methods of setting a fixed presentation level vs. performance intensity function
    • Subjects: listeners with hearing loss
    • Stimuli: NU6
    • Two fixed methods:
      • 40 dB SL re: SRT
      • Adaptive procedure (Levitt, 1978)
    • Results:
      • Adaptive procedure correctly estimated PB-max in 75% of cases
      • Fixed level of 40 dB SL re: SRT correctly estimated PB-max in 40% of cases
         

13

What do different hearing losses show in performance intensity functions?

  • a - normal hearing
  • b - conductive hearing loss
  • c - SNHL
  • d - SNHL 
  • e - SNHL - rollover, if significant - 8th nerve lesion
  • For c, d, e more gradual slope and lower PB max

A image thumb
14

Describe Guthrie and Mackersie's study on the comparison of presentation levels. What was their goal? Who were the subjects? What were the materials? What 4 methods of presentation level did they compare? What were their results? What was their conclusion?

  • Guthrie & Mackersie (2009)
    • Compared 4 methods of setting presentation level to yield maximum speech recognition for hearing-impaired listeners (aiming to avoid UCL)
    • Method: Compared 4 methods
      • Listeners: (n = 10/group, 4 groups)
        • gradually sloping losses: mild , moderate, mod-sev/sev
        • Sharply sloping losses
      • Speech materials: CASPA (computerized)
        • 10 CVC words/list with phoneme scoring - correctness of each phoneme
      • Obtained PI functions (SRT + 5 dB-> UCL –5dB)
        • UCL –5dB: gold standard
        • x dB SL re: SRT (20 –35 dB SL)
        • x dB SL re: 2kHz (10 –25 dB SL)
        • MCL 
    • Results:
      • L’s with mild losses: no difference in 4 methods
      • L’s with moderate losses: UCL –5dB> x dB SL re: SRT and MCL (no difference w/ x dB SL re 2k Hz θ)
      • L’s with severe losses: UCL-5dB > all other methods
      • L’s with steeply sloping losses: UCL-5dB > xdB SL re: SRT
    • Conclusion: method to set presentation level varies somewhat with degree and configuration of loss
      • Mild loss –any method is OK
      • Other degrees and configurations: vary somewhat
        • UCL –5 dB appears to be preferred in all degrees/configurations

15

How does word familiarization effect performance on word recognition? What are the implications of this?

  • Word Familiarity
    • individual’s lack of familiarity will affect score
    • Owens (1961)
      • Compared performance on two lists, matched phonetically but differing in word familiarity
      • Performance was significantly higher on familiar word lists
  • implications:
    • children: PB-K’s (6-12 yo), NU-CHIPS (3-5 yo)
    • foreign speakers: Spanish Lists;  Russian lists
       

16

Describe half lists vs. whole lists with word recognition. What are the advantages of half lists? What are the disadvantages of half lists? What should the decision be based on? 

  • Advantages of half lists
    • Shorter test time
    • May be sufficient for diagnostic purposes
    • Hi intra-list reliability (1/2 list to whole list)
  • disadvantages of half lists
    • Sacrifice phonetic balance (lose face validity)
    • Sacrifice inter-list equivalence
    • Lower reliability than full lists
    • More weight of random errors
  • decision based on purpose of testing
    • Estimating everyday performance -> full list
    • Diagnostic purposes -> half list may be OK

17

Describe binomial distribution and speech recognition scores. How is a significant difference determined?

  • Binomial Distribution and Speech Recognition Scores - what's a significant change in speech recognition scores? 
    • Thornton & Raffin (1978)
    • Uses the theorem to determine the s.d. associated with a particular mean score when the score is based on a binomial event (responses correct or incorrect, yes or no, right or wrong)
    • s.d. indicates the variability around the score; thus, a second score > this variability would be significantly different from the first score
      • What score would be significantly different from first score? If 2nd score differed by greater than s.d., significant difference 
         

18

What is the binomial probability theorem? What are its implications? What are its implications for clinical testing?

  • Binomial probability theorem
    • binomial distribution shows that the s.d. is inversely proportional to the square root of the number of test items (N)
      • s.d. will vary with percent correct score
    • implications:
      • as N increases -> s.d. decreases
        • Results are more reliable with longer list
      • as percent correct score varies                
        •  s.d. -> 1/( sqrt(n))
        • Scores near 0 and 100% are more reliable
        • Scores near 50% are least reliable
    • Implications for clinical testing:
      • Use longer lists when possible
      • Use tables to indicate s.d. of a given score to determine if a second score is significantly different from the first score
         

A image thumb
19

Describe Schlauch et al about word vs, phoneme scoring. What was the method (subjects, speech materials, conditions)? How were results analyzed? What were the results? What was the most effective test for improving precision of word rec tests?

  • Schlauchet al., 2014: evaluated measurement variability with word recognition performance using word vs. phoneme scoring (is better precision achieved w/ phoneme scoring?)
  • Method:
    • 24 NH adults
    • Speech material:  NU6
    • Conditions:  4 levels of speech-shaped noise to create 50 dB EM (twice), 52 dB EM, and 54 dB EM; speech level = 48 dB HL
      • EM - effective masking to simulate difficulty of hearing impaired individuals
    • Analyzed results per word (n = 50) and per phoneme (n = 150, but they are not independent)
  • Results:
    • Scores on half lists: higher variability than on full lists 
    • Phoneme scores were higher than word scores for same condition
    • SDs were smaller for phoneme scoring than word scoring
    • WRSs based on full lists analyzed by percentage of phonemes correct was the most effective test for improving precision of word rec tests

20

Describe the lexical neighborhood test. What was it developed for? What is a limitation of PB-K? What was the test based on? What were the vocab items? How difficult were the words? What is the LNT? What is the MLNT? Were the recordings male or female?

  • Developed for hrg-imp. kids using sensory aids
  • Limitation of  PB-K –vocabulary too difficult
  • test development:
    • based on NAM - neighborhood activation model  
      • When you hear a word you activate a lexical neighborhood 
        • Group of words that differ by one phoneme
          • Dense - cat hat bat pat
          • Sparse - juice
    • vocabulary items: taken from children’s speech (3-5 yrs)
    • lexical difficulty: half of the words are lexically easy and half are lexically hard - difficulty based on density of neighborhood and frequency of word
      • Lexically easy: used with high frequency and come from sparse neighborhoods (e.g. juice, good, six, cold, hard)
      • Lexically hard: not used often and come from dense neighborhoods (e.g., thumb, pie, wet)
    • LNT: 50-item lists of monosyllabic words
    • MLNT: 50-item lists of 2-or 3-syllable words
    • recordings using male talker

21

Describe the evaluation of the recorded version of the lexical neighborhood test. Who were the subjects? How many times were they ealuated? How were the stimuli presented? What were the results? What was the conclusion?

  • Evaluation of recorded version of test
    • 16 children with cochlear implants
    • evaluated twice
    • stimuli presented at 70 dB SPL in sound field
    • results:
      • Scores ranged from 4 –82% correct
      • lexically easy words> lexically hard words
      • high test-retest reliability
      • high interlist equivalence within each test
    • Conclusion: LNT and MLNT should yield estimates of spoken word recognition in kids 

22

Descibe the NAM in adults. What did Dirks et al examine? What were the results? What was the conclusion?

  • Dirks et al. study (2001)
    • Examined effects of word frequency, word density, and neighborhood frequency on word recognition performance in hrg imp. adults
      • Word frequency: frequency of occurrence of word in language
      • Word density: # of  phonetically similar words
      • Neighborhood frequency: frequency of occurrence of phonetically similar words
    • For normal and hearing impaired adults:
      • High frequency words > low frequency words
      • Words in low density neighborhoods > words in high density neighborhoods
      • Words in low frequency neighborhoods > words in high frequency neighborhoods
    • Conclusion: speech recognition tests should take into account the cognitive-lexical principles of NAM