Language Assessment in Natural Environments

1
Language Assessment in Natural Environments

• Judy K. Montgomery, PhD Chapman University, CA
• John Hansen, PhD University of Texas-Dallas
• Charles Greenwood, PhD University of Kansas
• D. Kim Oller, PhD University of Memphis
• ASHA Convention 2008 Chicago, IL

2

• This is an invited session of the 2008 Convention
  Committee on School Age Language.

3
Disclosure

• All four speakers serve on the Scientific
  Advisory Board for Infoture, a privately held
  company in Boulder Colorado, which developed
  LENA. They have no investment interests in this
  company.

4
Learning Objectives

• Participants will be able to
• 1. Recall measurement methods in natural
  environments
• 2. Identify speech recognition and recording
  systems
• 3. Analyze results of a study intended to
  replicate the findings of the classic Hart and
  Risley work
• 4. Interpret child language acquisition in a
  bilingual home

5
How we assess the language of very young children

• Formal, standardized tests
• Language Sample
• Parent reports
• Parent surveys and checklists
• Observations during play in therapy
• Predict what occurs in the natural environment
  (without the presence of the clinician)

6
These approaches are all labor intensive,
tedious, and in the end, still quite subjective.

• We need to
• have access to better assessment tools.
• be focused on working with caregivers
• measure change
• give meaningful feedback to caregivers
• take advantage of natural environments

7
We would like to know if

• Parents talk to their children?
• Clinical treatment underway is being carried out
  in other settings?
• Children are engaged in a richly nuanced language
  environment?
• We can identify early signs of language delay,
  even autism, within the acoustic speech
  signal? (Warren, 2008)

8
Natural Environments

• Settings that are typical for infants/toddlers
  without disabilities.
• Homes, early care programs
• Community settings where families spend time with
  their children
• (IDEA, 2004 Woods, 2008)

9
Natural Environments, cont.

• More than a location for services
• Familys typical and valued activities
• Parents as partners
• Communication intervention is integrated into the
  childs daily routines
• (Woods, 2008)

10
Natural Environments, cont.

• A process, not a place
• Role release
• Focus on interaction between caregiver and child,
  not delivering services directly to child
• Embed into existing routines
• (ASHA, 2008b)

11
Why do we prefer natural environments?

• ASHA document (2008) policy
• Evidence-based practice
• Relationship building
• Authentic

12
Language Environment Analysis

• Is there another way to collect the data to
  assess language in natural environments?
• LENA- An automatic speech monitoring and analysis
  tool.

13
LENA

• Child wears 2-ounce digital recorder
• Records up to 15 hours of child and adult
  vocalizations
• Uploads digital speech files with USB port
• Analyzes number of adult words, child words,
  conversational turns.

14
Measurement drives science

• Measurement methods and technology shape,
  enable, and constrain advances in all sciences
  (p. 2)
• LENA may represent a transformative
  technological breakthrough in the measurement of
  human behavior that is perhaps without parallel
  (p. 5)
• (Warren, in press).

15
New way to assess-- and intervene

• Use the standardized tools you trust.
• Observe child in therapy environment.
• Take the LENA home and use it for three days.
• Review the data with the parents and see if they
  recommend some changes.
• Continue interventions.
• Assess again with LENA in 4 weeks.

16
Replication of Hart Risley findings (1995)

• Based on seminal study conducted in 1970s,
  published in 1995.
• Recorded 1 hour of talk in the homes for 42
  children from birth to age 3.
• Thousands of hours ( 7 years of transcriptions)
  to produce results.
• Can now finally be replicated.

17
Talk Time

• Assess in the natural environment
• Collect what adults and child say
• Monitor change over time
• Add an objective element to the evaluation process

18
Talk Time

• Siblings should read aloud for their own practice
  and to increase word count for the baby.
• Recording the number of words spoken to child at
  home all day long.

19
How could you use this measurement tool?

• Collect authentic home interaction data before
  child is assessed.
• Assess at-risk students
• Monitor a parent program
• Monitor effects of intervention
• Measure effects of objective feedback- do
  caregivers change their habits?

20
More ways

• Measure talking environment of students with
  cochlear implants
• Use of AAC at home?
• Language development in Autism Spectrum Disorders
• Determine amount of TV time in home

21
How did we use this tool?

• Collect Adult Word Counts (AWC)
• Collect Child Word Counts (CWC)
• Collect Conversational Turns (CT)
• Young children with and without known
  disabilities.
• Do children with disabilities get the same rich
  language environment?

22
Preliminary findings from the pilot study

• SLPs used it with their own children.
• Was taxing to say 20,000 words a day to child.
• Found that book reading was a savior!
• More words in an environment with many adults and
  siblings
• Ever-changing environments provided a chance for
  many more words each day
• First born get more of these high AWC days

23
Preliminary confirmationsfrom the pilot study

• Parents estimated they talked more with their
  children than they actually did
• Mothers did most of the language training
• Most CTs occurred in the late afternoon
• Parental confidence increased after reviewing the
  data
• (Shubin, 2008)

24
Evidence-based practice for

• Prevention- at risk children
• Assessment- whats happening at home
• Intervention- will parents increase their talk if
  they see the word counts

25
What did parents tell us?

• The biggest benefit of LENA is seeing how much
  you really talk. I thought I talked much more
  than I really do.
• I like being able to tell where the gaps in our
  day are and where we can improve.

26
What did parents say?

• I love to see Erik improve from the 80th to the
  90th percentile in conversational turns.
• It seems to be an early warning device for
  language development.

27
Meet Raja

• Actual data
• 832 conversational turns in 10 hours
• 98th percentile
• 18,116 adult words in 10 hours
• 92nd percentile

28
The childs first word is often marked with
photos, video or audio recordings for the baby
book, phone calls to grandparents (Woods, 2008,
p. 14).

• ,

29

• No wonder those at the leading edge of literacy
  want to increase the quantity and quality of
  conversations between parents and children
  beginning at birth.
• (Pappano, Harvard Education Letter, May/June,
  2008, p. 1)

30
References

• American Speech Language Hearing Association.
  (2008b). Roles and responsibilities of speech
  language pathologists in early intervention
  Position Statement. Rockville, MD Author.
• Gilkerson, J. (2008). The power of talk. Council
  of State Association Presidents, Saratoga, NY.
• Hart, B. Risely, T.R. (1995). Meaningful
  differences in the everyday experiences of young
  American children. Baltimore Brookes.
• Pappano, L. (2008). The power of family
  conversation. Harvard Education Letter, May/June.
• Shubin, J. (2008). Early intervention in natural
  language environments. Unpublished study. Chapman
  University, Orange, CA.
• Warren, S. F. (in press). Measurement and the
  future of behavioral science. Psychology in
  Intellectual and Developmental Disabilities, 34.
• Woods, J. (2008). Providing early intervention
  services in natural environments. ASHA Leader,
  March 25, 2008 14-17.

31
Robust Speech Processing
Is there a way to measure the
words a child hears?
John H.L. Hansen
Center for Robust Speech Systems (CRSS) Erik
Jonsson School of Engineering Computer
Science Department of Electrical
Engineering School of Brain Behavioral Sciences
(Speech Hearing) University of Texas at
Dallas Richardson, Texas 75083-0688, U.S.A.
ASHA Convention, Nov. 20-22, 2008
32
OUTLINE

• GOALS Speech Recognition (Adult-Child)
• Challenges for Speech Recognition
• Structure of Speech Recognition Engines
• Features, Acoustic Models, Language Models
• Counting Words a Child Hears
• Assessing Stress in Adult-Child Interactions

33
INTRODUCTION

• PROBLEMS / GOALS
• Assessing Child Language Development can
  be much more effective when naturalistic
  data is available from in-home
  environments.
  Is it possible to obtain such naturalistic
  data?
• Hart Risley Study (1995) variation in language
  ability is relative to the amount parents speak
  to their children
• What Language Technology Challenges Exist?
• Is it possible to count the number of words a
  child hears using speech recognition?
• Is it possible to assess Adult-Child interactions
  to determine if stress impacts adult-word-count
  (AWC) or adult-child conversational turns (CT)?

34
Speech Recognition Challenges
Why Speech Recognizers Break!

• Speaker Based Problems

• Context Based Effects

• Stress Emotion
• Lombard Effect / Noise
• Psychological Task Demands
• Accent / Language / Dialect
• Speaker Differences (Age, Gender, Vocal Tract)
• Speaking Format/Structure
• Read 2-way Conversational Spontaneous,
  Monologue, Dialog Response Telephone

• Homonyms (English 10,000 Japanese 120)
• Confusable
• (Take, Stake, Straight Cake, Kate)
• Ambiguous
• Jeet Yet? Its ours vs. It sours
• Nice guys vs. Nice skies
• Its hard to recognize speech vs.
• Its hard to wreck a nice beach

Um, I just wanna, I just want to say, I
dont know what I want to say.

• Communication Based

• Environmental Based

• Microphone
• Voice Compression
• Channel / Mobile Cellular

• Acoustic Noise
• Room Reverberation
• Physical Task Demands

35
Speech Recognition in Audio Streams
What are the differences?

• Speech Recognition LVCSR
• Keyword Spotting
• Topic Spotting
• Word Count Conversational Turn Estimation

// speech (1 or more speakers)
joe ate his soup. this field of beets
is ripe and ready. orange juice tastes
funny

• recognize every word

// speech (1 or more speakers)
soup beets
orange juice

• given isolated word search list recognize
  keywords, ignore the rest

// speech (1 or more speakers)

• given search topic recognize topic words if
  hit rate reaches threshold, spot topic

// speech (1 or more Adult speakers, child,
Noise TV, radio, distant speakers)
Adult Child Noise

• do not recognize words only estimate count and
  turn taking remove noise in count

36
Speech Waveform Content
(adapted from J. Campbell, et al.s slide
presentation on Speaker Recognition- ICASSP-03)
37
Speech Recognition Benchmarks
DARPA Speech Recognition Benchmarks (1988-99)
None of these have focused on Naturalistic
Data None focus on Adult-Child interaction All
generally have defined vocabulary
space unrestricted data/topic/vocabulary not
typical
38
Flow Chart of ASR Process
Speech Recognition
Acoustic Model
Lexicon
Adaptation
Audio sampled at 8kHz or 16kHz
Feature Extraction
Pattern Matching / Search
Speech Detection
Recognized Words
Language Model
Adaptation
I wanna fly to Chicago tomorrow
39
Speech Recognition
ASR Front-End Processing
Speech
Spectrogram
Observation Feature Sequence

O3
O4
O5
O1
O2
O6
O7
40
Speech Recognition
Anatomy of an Automatic Speech Recognizer (ASR)
Lexicon(phoneme-based)
Language Model(n-gram)
Feature Extraction (spectral analysis)
Text Timing Confidence
Optimization (Viterbi Search)
Speech
Acoustic Model(HMM)
MFCCs
41
Speech Recognition
ASR Speech Feature Extraction MFCC
Mel-Frequency Cepstral Coefficients (MFCC)
42
Speech Recognition
ASR Hidden Markov Model Acoustic Modeling
43
ASR Problem Formulation
Speech Recognition

• Given a sequence of observations extracted from
  an audio signal,
• Determine the underlying word sequence,
• Optimize

Probability of Word Sequence
44
Understanding Speech Recognition
Speech Recognition
Word Sequence
45
Lexicon
Speech Recognition

• Encodes knowledge of phonetics, word
  pronunciations
• Defined by a symbol set for each language
• CRSS-UTD uses CMU Sphinx-II symbol set 51
  symbols (including silence)
• Examples
• Friday F R AY D EY -or- F R AY D IY
• Communicator K AX M Y UW N AX K EY DX AXR
• For F AO R -or- F AXR -or- F R AXR
• Freely available US English lexicon from CMU
• http//www.speech.cs.cmu.edu/cgi-bin/cmudict

46
Statistical Language Modeling
Speech Recognition
47
Speech Recognition
HMMs
MFCCs
Bi-Gram
Word list with Pronunciations
Recognized Word Sequence
48
Counting Words a Child Hears

• Need to capture Naturalistic Adult-Child
  conversations

• Speech Recognition to count words, would need
  complete dictionary listing in Lexicon (300,000
  words)

• Instead of DIRECT ASR, (i) segment data, (ii)
  classify into broad classes Key Child, Clear
  Adult, TV, Others, (iii) perform Phone
  recognition on Clear Adult with least squares
  linear regression

49
LENA Speech System Layout
50
AWC System Performance

• Performance Relative root mean square error
  (RMSE)
• Overall relative error 2 RMSE reduces over
  time

51
Infoture Database

• Infoture, Inc. database consists of adult-child
  interactions in a natural home environments.
• Details
• Over 3000, 12-16 hour recording sessions from
  over 460 families (totaling over 61,000 hours of
  recordings),
• Corpus is a balanced for child gender
  (male/female) age (2-36 months).
• The participating families are from all
  socio-economic status.

52
STRESS in Adult-Child Speech

• Focus Study focuses on adult-child interactions
  in real-life scenarios to
• Build a reliable stress detection system,
• Relate adults stress to a adult word count (AWC)
  and conversational turns (CT) two indicators of
  childs language development.

• Adult stress state is defined into two
  categories
• Neutral State includes plain interaction
  (interaction void of any stress), joy, happy
  state of emotion
• Stress State includes all negative emotions
  anger, frustration, sadness.

53
Core Database

• Due to IRB requirements, we used core database
  from 20 families, 12 hours per family, totaling
  320 hours of data.
• Utterance pool greater than 2 secs includes 2193
  utterances, perceptual stress assessment to
  categorize stress speech (SX), neutral speech
  (NX), undecided (UX).

85.62 of the conversations lt 2 sec.
(adult-child interactions shorter durations)
Perceptual stress assessment
54
Stress Detection Accuracy

• Female adult speech was correctly classified as
  either neutral or stress speech with 72.28
  accuracy, male adult speech was correctly
  classified with 67.53 accuracy.

Stress detection for female higher vs. male speech
55
Relating Stress Detection Accuracy with Child
language development metrics

• Adult speech for females and males
• Female speech shows larger variation (s 0.4327)
  under neutral indicating that mothers used
  motherese style, while stress speech shows less
  variations (s 0.0561).
• Male speech shows limited variation while
  interacting with child under neutral conditions,
  indicating males do not change their speech as
  much while speaking to their child.

56
Conclusion Discussion

• Discussed Challenges for Automatic Speech
  Recognition for Naturalistic data
• Anatomy of a Speech Recognition Engine
• (feature extraction, lexicon, Acoustic Model
  recognition, Language Model recognition)
• No ASR Systems available for Adult-Child
  interactions possible to develop alternative
  method to estimate Adult-Word-Count and
  Conversational Turns, WITHOUT lexicon
• LENA System Employed for Assessing impact of
  Stress
• Can Detect Adult Speech Under Neutral vs.
  Stress 70.2
• correlation of adult stress with ratio Utterance
  Length / Adult Word Count is more prominent for
  male adults compared to female.

A Window into the Adult-Child Language
Environment
Combination of hardware software of LENA
provides a practical tool for Speech Language
Pathologists and other professionals to assess
child language development
57
References

• 1 S. Barnes, et al, Characteristics of Adult
  Speech which Predict Childrens Language
  Development, J.
• Child Language, (10)65-84, 1983.
• 2 B. Hart an T. Risley, Meaningful Differences
  in the Everyday Experiences of Young American
  Children,
• Baltimore, Maryland, Brookers
  Publication, 1995.
• 3 J.H.L. Hansen, Analysis and Compensation of
  Speech Under Stress and Noise For Environmental
• Robustness in Speech Recognition,
  Speech Communication, (20)151-173, Nov. 1996.
• 4 J.H.L. Hansen,et al, "The Impact of Speech
  Under Stress' on Military Speech Technology,"
  NATO
• Research Technology Organization RTO-TR-10,
  AC/323(IST)TP/5 IST/TG-01,
• March 2000 (ISBN 92-837-1027-4).
• Stress Classification
• 5 G. Zhou, J.H.L. Hansen, and J.F. Kaiser,
  "Nonlinear Feature Based Classification of Speech
  under Stress," IEEE Trans. Speech Audio Proc.,
  vol. 9, no. 2, pp. 201-219, March 2001.
• 6 D. Cairns, J.H.L. Hansen, "Nonlinear Analysis
  and Detection of Speech Under Stressed
  Conditions," Journal Acoustical Society America,
  96(6)3392-3400, December 1994.
• 7 B.D. Womack, J.H.L. Hansen, "Classification
  of Speech Under Stress using Target Driven
  Features," Speech Communications, Special Issue
  on Speech Under Stress, vol. 20(2), pp. 131-150,
  Nov. 1996.
• Robust Speech Recognition in Stress Noise
• 8 S.E. Bou-Ghazale, J.H.L. Hansen, "A
  Comparative Study of Traditional and Newly
  Proposed Features for Recognition of Speech Under
  Stress," IEEE Trans. Speech Audio Proc., 8(4)
  429-442, July 2000.
• 9 R. Sarikaya, J.H.L. Hansen "High Resolution
  Speech Feature Parametrization for Monophone
  Based Stressed Speech Recognition," IEEE Signal
  Proc. Letters, vol. 7, no. 7, pp. 182-185, July
  2000.
• 10 B.D. Womack, J.H.L. Hansen, "N-Channel
  Hidden Markov Models for Combined Stress Speech
  Classification and Recognition," IEEE Trans.
  Speech Audio Proc., 7(6)668-677, Nov. 1999.
• 11 S. E. Bou-Ghazale, J.H.L. Hansen, "Stress
  Perturbation of Neutral Speech for Synthesis
  based on Hidden Markov Models," IEEE Trans.
  Speech Audio Proc., 6(3)201-216, May 1998

58
Community-based Exploration of the Home Language
Environment Using Automatic Speech Processing

• Charles R. Greenwood, Kathy Thiemann, and the
  Early Speech Project Workgroup
• Juniper Gardens Childrens Project
• University of Kansas
• November, 2008

59
Their Goal?

• Improve the developmental experiences of area
  children, thereby
• Improving the academic and social achievements of
  the children
• The mission of the Juniper Gardens Childrens
  Project in Kansas City (www.jgcp.ku.edu )

60
What they Discovered?

• Vast differences in childrens vocabulary growth
  and other language indicators

61

• Vast differences in the home language environment
  (words addressed to the children)

62
Limitations?

• HR findings were based on 1-hr recordings of
  daily prime time per month
• We dont know how representative these 1-hr
  estimates were?
• We dont know the patterns of talk over an
  entire day
• What aspects of HR findings can be replicated?

63
Aims

• To conduct a more time intensive, longitudinal
  study of childrens language learning using the
  HR framework
• To explore differences in childrens home
  language environment and childrens talk
• To replicate aspects of HR findings

64
Research Questions

• For Adult Word (AW), Child Vocalization (CV), and
  Conversational Turn (CT) frequencies,
• What was the overall mean and variation? Do
  families show individual differences?
• How comparable were estimates to HR?
• What do the over time trajectories look like
• over months within one 12-hr day
• Was there a relationship between AW, CV, CT
  other commonly used criterion measures?

65
Design

• Longitudinal, age-at-start cross-sectional,
  repeated measures design
• Study Duration 10 months
• Setting Childrens homes and community

66
Sample

• A sample of middle-SES English speaking families
  with typically developing children between 12 and
  20 months of age was recruited.
• Families were recruited from Metro Kansas City
  communities (e.g., home visiting programs,
  mailings to families, advertisements, and
  postings.

67
Sample

• Enrolled Sample (N 39)
• Analysis Sample (N 37)
• Age at Start (M 15.6 mos, SD 3.3)
• Race/Ethnicity
• 78 White, 10 Black, 7 Hispanic,
• 3 Pacific Islander, and 3 multi-race)
• Primary Caregivers
• Age (M 31.2 yrs, SD 6.2)
• Education (All High School/GED or above)

68
Measurement

• A prototype version of the LENA System was used
  by parents in the home (http//www.lenababy.com/Le
  naSystem/PowerOfTalk.aspx )
• A small digital audio recorder placed in the
  pocket of the childs LENA clothing was used. It
  was turned on by the parent to make a 12-hour
  automated recording (see illustrations).

69
Processing Accuracy
70
Procedures

• LENA automated processing software was used to
  extract three HR speech-related vocalization
  scores nested within children-families, nested
  within recording dates
• These were
• AW Frequency of Adult Words Addressed to the
  Child (those close enough to be heard by the
  child)
• CV Frequency of Child Vocalizations
  (speech-related utterances)
• (CT) Frequency of Conversational Turns (changes
  in floor holding between child and adult)

71
Procedures (continued)

• All participants received remuneration for making
  and returning digital home recordings at the rate
  of 25 per recording
• The total number of 12-hour digital recordings
  available for use in this report was 720 (8,640
  hours) 

72
Number of 12 hr Recordings per Month of Age
680 recordings in this analysis
73
Criterion Validity Measures

• Bayley Scales of Infant Development (BSID-III)
• Preschool Language Scale (PLS-4)

74
Results

• For Adult Word, Child Vocalization, and
  Conversational Turn Counts,
• What was the overall mean and variation? Do
  families show individual differences?

75
Overall Descriptives
76
Results

• For Adult Word, Child Vocalization, and
  Conversational Turn Counts,
• What do the over time trajectories look like
• within one 12 hr day
• over months

77
Circadian Rhythms of Daily Talk
78
Patterns of Daily Talk by Age
79
Words Heard by the Child (AW)
80
Child Vocalizations (CV)
81
Conversational Turns (CT)
82
Individual Family Differences
83
Results

• For Adult Words
• How comparable were estimates to HR?

84
Words Heard per Hour by Age
85
Cumulative
10.8 Million
86
Was there a relationship with other criterion
measures?

• Adult Word (AW) vs.
• Bayley Cognitive (.29)
• PLS Receptive (.39)
• PLS Expressive (.11)
• Total Language (.29)
• Conversational Turns (CT) vs.
• Bayley Cognitive (.51)
• PLS Receptive (.38),
• PLS Expressive (.30)
• Total Language (.37).

• Child Vocalizations (CV) vs.
• Bayley Cognitive (.49)
• PLS Receptive (.14)
• PLS Expressive (.15)
• Total Language (.16).

87
Summary

• Automatic speech processing (ASP) was used to
  extend HRs findings in terms of greater sampling
• Total hours per day per month
• Multiple recordings per week per month
• 12 vs. 1 Hour within a day
• Unlike HR, ASP does not yet achieve automatic,
  vocabulary recognition, rather speech-related
  vocalization counts and turns
• Like HR, vast differences were found in families
  patterns of talk in the home
• LENA indicators were positive correlates of
  cognitive and language measures

88
In terms of construct validity relative to
theory,

• Sizeable individual family differences in the
  home language environment were demonstrated
• A higher proportion of adult language heard by
  the child was contributed by female, rather than
  male caregivers
• Child vocalizations and conversational turns
  increased with age
• Hourly patterns of variation in the home language
  environment were linked to sleep, prime time, and
  feeding periods of the day as would be expected.

89
Implications for Future Research

• Replication of a number of HR findings
  economically in much larger samples is now
  possible
• The hypothesis that the home language environment
  drives language learning is open to investigation
• Developmental as well as intervention studies are
  possible in a range of important areas

90
Evaluating early child language input and
language learning New tools for naturalistic
sampling and assessment

• D. Kimbrough Oller
• The University of Memphis, USA
• The Konrad Lorenz Institute for Evolution and
  Cognition Research
• Altenberg, Austria

91
A fundamental multilingualism question

• A standard generative linguistics assumption
  Input, no qualifications, drives the way language
  develops
• And presumably drives the choice of language in
  multilingual circumstances
• A contrasting functionalist assumption
• Input directed toward the child is more
  significant than other input in the environment
• Both in how language structure emerges and in
  language choice
• This is a question that has roots in the poverty
  of the stimulus argument and in critiques of the
  argument
• Many implications for how language is learned
  generally
• And regarding clinical recommendations for
  multilingual families

92
An empirical evaluation of the role of input
amounts in various languages

• The availability of new hardware and software
  makes it possible to address these questions with
  new empirical power
• The key issues are naturalistic recording
• And representative sampling from those recordings

93
Three languages in this childs life

• Mother, Austrian, spoke German to the child very
  predominantly
• Father, American, spoke German (as well as
  possible) also very predominantly
• Governess, Central American, spoke Spanish almost
  exclusively
• Mother and Father tended to speak English to each
  other, and the broader environment was very
  predominantly English

94
How can the recordings be used in practical
applications?

• There is no reliable automatic identification
  even of adult speech in naturalistic environments
• Much less of infant and child words
• It takes substantial resources to transcribe from
  recordings to identify the words spoken by both
  adults and children
• Why would we wish to record so much if we still
  have to listen to extremely lengthy recordings

95
The new technology

• LENA software counts adult words and infant and
  child vocalizations
• This it does quite reliably
• This information is available in the form of a
  handy interface that makes it possible to locate
  high density or low density periods of vocal
  interaction
• With foresight, one can select a small sample to
  transcribe
• The sample can be specifically chosen for
  representativeness across a wide range of
  recording circumstances

96
The present study

• 11 recordings averaging 10 hours for each of the
  11 days
• Encompassing the great bulk of the childs waking
  hours
• 13-month period from 11 to 24 months of age
• Days for recording had been chosen to include
  representation for all the languages of input
• Selections of 39 5-minute samples of recording
  were then made, representing high volubility
  periods as indicated in the LENA reports from
  across the 11 recording days

97
Balancing the sample of input distribution by
language, 1

• A key to the design of this approach is to
  characterize the contact hours proportionally for
  the major participants in input (Mother, Father,
  Governess)
• And further to estimate the amount of Other
  circumstances of input
• Quantitative records had been kept of the
  governesss presence in the household
• And a quick review of recordings made in the home
  on those days confirmed the parents report that
  the governess was usually alone with the child
  when she was present
• As well as estimates of how much time Mother and
  Father spent alone and together with the child

98
Balancing the sample of input distribution by
language, 2

• Proportions of contact across the 13-month period
  were then calculated for
• Mother and Father with the child 28.5
• Mother alone with the child 27.6
• Father alone with the child 15.6
• Governess alone with the child 15.5
• Other circumstances 12.7
• Data could then be sampled from each of these
  conditions and comparisons of input could be made
  in terms of this distribution of input

99
Listening illustration and Reliability check

• Here we use LENA software to play a segment of
  recording and illustrate the counting method
• Reliability checked on 7 of the 39 samples (28
  word-count values)
• Correlation of counts 0.97
• Average discrepancy of 3.9 words across a range
  of 0 to 93 words per word-count value (0.23 of
  mean word-count value)

100
Observed raw word counts (tokens)
101
Observed proportional word counts (tokens)
adjusted for input time distribution
102
Observed proportional word counts (tokens)
adjusted for input time distribution by language
103
Observed proportional word counts
(tokens)adjusted for input time distribution by
language
104
Observed proportion of lexical types from the
child compared with proportional input amounts
105
Conclusions, 1

• This method allows very quick access to
  significant aspects of the naturalistic
  environment even in the absence of automated
  recognition of words
• We learned that (consistent with the
  functionalist view) input effect depends on
  directedness toward the child
• Having adjusted for the total amount of input,
  and considering word types, it was
• 14 times more likely that a Spanish word would be
  represented in the observed lexicon than an
  English one
• And 6 times more likely that a German word would
  be represented

106
Conclusions, 2

• Considering word tokens, after adjustment for
  input time distribution, it was
• 7 times more likely that a Spanish word would
  occur than an English one
• And 5 times more likely that a German word would
  occur than an English one
• Why? The pattern is clearly associated with
  directedness of input toward the child
• Directed input words were almost 10 times more
  likely to be German than English
• And 2.5 times more likely to be Spanish than
  English

107
Conclusions, 3

• But undirected English words were 10 times more
  frequent than Spanish words
• And 2.3 times more frequent than German words
• Perhaps, then, the most notable finding was the
  underrepresentation of the most frequent overall
  undirected input type, in this case, English
• So the type of input really does matter in a big
  way whether speech is directed toward the child
  plays a huge role in what language the child
  learns to use

108
Conclusions, 4

• This outcome should be no surprise, in my opinion
• But at the same time we have demonstrated the
  effects quantitatively
• And illustrated that naturalistic sampling using
  all-day recordings can now provide a platform for
  acquiring new quantitative perspectives on
  language acquisition
• The new tools make it possible for us to obtain
  representative samples at low cost

109
Final Thoughts

• The present study is only one example of how we
  have entered a new era of research on language
  development
• My research agenda has been radically altered now
  that naturalistic sampling is available
• Automated analyses are rapidly being developed to
  eliminate or dramatically reduce the costs of
  acquiring quantitative information
• Clinical implications are extraordinary
• For screening and diagnosis
• For providing clinicians with a view into the
  home
• For supplementing therapy at the clinic with
  semi-supervised therapy conducted in the home and
  monitored by the clinician