Speech Recognition Technology Applications

							Speech Recognition
Technology Applications



 Denise Bilyeu, M.S. CCC-SLP
 Scottish Rite Computer
 Supported Literacy Program
 Munroe-Meyer Institute
 Omaha, NE
Speech Recognition

Utilizes hardware and software to
 transcribe spoken words into orthographic
 text
Allows users hands free operation of
 computer systems
Applications for Persons
with Disabilities

Academic opportunities
Vocational opportunities
Access to WWW
Implementation Issues

System Training Requirements
Dictation in Written Form
Absence of Graphical Representation
Functional Grade Level
Dictation Environment
Higher Order Organizational
 Skills/Strategies
System Training
Requirements

Samples of training protocol text (500
 words) were taken from each of the
 following programs
  Dragon Naturally Speaking Standard*
  Dragon Naturally Speaking Teen*
  IBM Via Voice Gold
  L & H Voice XPress

  *two samples were analyzed and averaged
Samples were analyzed using Readability
 Stack (Tice, B. 1990)
  Flesch Index
  Dale Index
  Dale-Chall Formula
  Fry Readability Graph
Flesch Index

(RE= 206.835 - (1.015 x words/sentence) - (84.6
 x syllables/word)
Rates text on a 100 point scale
High scores indicate easier reading levels
Reading Ease based upon
  Mean Sentence Length
  Syllables per 100 words
Dale Index

DI = 11.534 - (.053 x RE)
Based on the Flesch Index Reading Ease
 Score
Dale-Chall Formula

Reading Grade Score (RGS) = .1579 x DS
 (Dale Score) + .0496 x SL (Sentence
 Length) + 3.6365
  Dale Score = % of words not on Dale list of
   3000
  Sentence Length = average # of words per
   sentence
Fry Readability Graph

Yields Readability Grade Score (RGS)
 based upon:
  Syllables per 100 words
  Sentences per 100 words


Average the RGS for 3+ random passages
 for reliable score
Dale-Chall Analysis

                                       Mean Sentence   Reading Grade
                                          Length           Score

  Dragon Naturally Speaking Standard       17.43           8.60
  Dragon Naturally Speaking Teen           14.18           8.47
  IBM Via Voice Gold                       10.44           7.46
  L & H Voice XPress                       12.39           7.70
Fry Readability

                                     Sentences /   Grade Level
                                      100 words

Dragon Naturally Speaking Standard       5.8           8
Dragon Naturally Speaking Teen           7.3           6.5
IBM Via Voice Gold                       9.6           4
L & H Voice XPress                       8.1           7
Flesch Index

                                     Flesch Index


Dragon Naturally Speaking Standard      68.59
Dragon Naturally Speaking Teen          78.24
IBM Via Voice Gold                      86.26
L & H Voice XPress                      74.13
Dale Index

                                     Dale Index


Dragon Naturally Speaking Standard     7.90
Dragon Naturally Speaking Teen         7.39
IBM Via Voice Gold                     6.96
L & H Voice XPress                     7.61
Conclusions

4th grade minimum literacy level required
 to train voice recognition programs (most
 programs need 6th to 8th grade reading
 levels)
Respiratory support sufficient to produce
 sentences of M = 10.44 words
No statistically significant differences in
 training protocols
Dictation in Written Form

Dictation vs. Conversational speech
Children produce 86% more words in
 slow dictation than in writing and 163%
 more words in normal dictation than in
 writing (Breeder & Scardamalia)
  Process is vastly different
  Dictation skills must be taught
Absence of Graphical
Representation

Difficulty with dictation is often attributed
 to absence of graphical representation;
 may cause problems in text development
 and revision (Wetzel)
Speech Recognition has graphical
 representation, but often with a delay that
 interrupts the dictation process
Functional Grade Level

Classroom placement and curriculum
 demands contribute to written text needs
Written text requirements may not be
 extensive enough to warrant a Speech
 Recognition system
Consider cognitive and/or language skills
Dictating Environment
Voice recognition requires an environment
 relatively free of auditory stimuli
Ambient noise will effect the system’s
 ability to function well
Dictating may be disruptive to others
Removal from the environment may solve
 dictation problems, but result in
 educational or vocational disruptions
Higher Order Organizational
Skills / Strategies

Persons must have cognitive abilities to
 dictate and often need strategies to help
 with the process
Pre-Writing Strategies
  Writing instruction
  Planning
  Outlining/Mapping
     Inspiration
Evaluation

Intelligibility
  Sentence Intelligibility Test                (Yorkston, Beukelman & Tice,
    1991)

     Utilizes ten unrelated sentences
     Transcribed by unfamiliar listeners
     Variables elicited
            • Intelligibility (% of intelligible speech to unfamiliar
              listener without context)
            • Rate of speech

Grade/Literacy Level
Fluency of Dictation
Attention to task
Writing/Dictating environment
Trial with voice recognition system
  set up microphone/sound system to see if
   voice is perceived
  run system training session if user is capable
  dictate known passages that require little
   cognitive demand e.g., pledge of allegiance
  dictate text that requires cognitive demand,
   short expository
Alternate means for training systems
  Utilize another person with similar voice
   characteristics
  Transcribe training protocols and allow user
   to learn and practice dictating
  Transcribe training protocols and dictate to
   tape for user to listen to while dictating
Case Studies
Janae

9 years old
Athetoid Cerebral Palsy
Sentence Intelligibility Test Score - 10%
Current System
  Discover Board, Mouse Key
Reason for Referral
  Mousing slow and fatiguing
Evaluation Tool
  Dragon Dictate v. 3.0


Evaluation Results
  With no training, could utilize Mouse Grid
   with 80% accuracy, after one hour session,
   could utilize Mouse Grid with 95% accuracy.
  With extensive training could dictate small
   amounts of text
Voice Recognition Status
  Utilizing Dragon Dictate Mouse Grid on trial
   basis
  Training on selected, commonly used words
   in progress to determine efficiency and
   fatigue effect of dictating text
James
14 years old
Learning Disabled, reading and writing
 skills 4 years below grade level
Sentence Intelligibility Test score = 100%
Reason for Referral
  Slow input method
  Input impeded cognitive writing process
  Inability to monitor written work
Evaluation Tool
  Dragon Naturally Speaking Standard
  Dragon Naturally Speaking Teen
Evaluation Results
  Training materials printed and practiced
   before actual program training
  Training required 2 weeks, 3 sessions/week
  Needed alternate text program to review text
  Worked on phrasing, assisted punctuation
Voice Recognition Status
  Uses voice recognition at home for
   homework and correspondence
  Does not use voice recognition at school
Brett

18 years old
Quadriplegia, ventilator dependent
Sentence Intelligibility Test score = 100%
Current system
  EZKeys for Windows with Morse Code input
   via pneumatic switch
Reason for referral
  slow input method
Evaluation Tool
  Kurzweil
Evaluation Results
  Ventilator had to be physically blocked at
   Brett’s neck and in the back of the wheelchair
  Training on segmentation of words and
   phrases was necessary
  Training required one month
Voice Recognition Status
  Able to use voice recognition at home for
   homework, correspondence and the internet
  Unable to use voice recognition at school
   because of ambient noise and disruption that
   dictation causes
Katie

16 years old
Traumatic brain injury
Sentence Intelligibility Test score = 89%
Current system
  regular keyboard with track ball
Reason for referral
  slow input method
  fine motor movement fatiguing
Evaluation tool
  Dragon Naturally Speaking Standard
Evaluation results
  Unable to train system during evaluation
   likely because of nasal emission on specific
   sounds, effecting the intelligibility of
   surrounding sounds
  Palatal lift was fitted subsequent to
   evaluation, but further voice recognition
   evaluation was not done
Voice Recognition Status
  Unable to utilize voice recognition at time of
   evaluation
  Further evaluation was not done as fine
   motor abilities were improving and alternate
   strategies (word prediction, abbreviation-
   expansion) were effective
John
53 years old
Friedrich’s Ataxia
Sentence Intelligibility Test score = 53%
Current system
  EZKeys for Windows scanning via pneumatic
   switch
Reason for referral
  slow input method
  alternate access for versatility and fatigue
Evaluation Tool
  Dragon Naturally Speaking Standard
  IBM Via Voice Gold
Evaluation Results
  Unable to train system after extensive trial
   period (4 weeks, daily)
  System would not “perceive” John’s voice
Voice Recognition Status
  Unable to utilize voice recognition
  Trial with Dragon Dictate scheduled
Clinical Implications

Decrease intelligibility results in decreased
 success with voice recognition
Intelligibility may NOT predict success
 with voice recognition
Rate of speech may effect success with
 voice recognition
Future Directions

New voice recognition programs require
 minimal training
New programs that do not learn as they
 are used are in development
New programs that utilize a standard set
 of distinct “sounds” are in development