The TBALL Project Data Collection: Making a Young Children's Speech Corpus

1
The TBALL Project Data Collection Making a
Young Children's Speech Corpus Abe Kazemzadeh,
Hong You, Markus Iseli, Barbara Jones,
Xiaodong Cui, Margaret Heritage, Patti Price,
Elaine Anderson, Shrikanth Narayanan, and Abeer
Alwan University of Southern California,
University of California Los Angeles, and
PPRICE Speech and Language Techology
Results and Observations

• Data Collection Motivation
• Establish a corpus for studying child and
  non-native speech.
• Build speech applications for under-represented
  populations.
• Analyze pronunciation variation.
• Test bed for our target child-computer
  interface.
• Test hardware, animations, timing, vocabulary,
  etc.
• Measure children's ability with respect to grade
  level and other factors

• Project Goals
• Automation of literacy assessment measures
  using speech and language technology.
• Development of standards and methods for
  reliable, objective assessment.
• One-on-one interaction with child, which leaves
  teachers with more time for teaching.
• Focused on fair assessment robust to dialect
  variation including nonnative speakers.
• Support for teacher feedback and database
  records.

Native Language Distribution of Recorded Subjects
Transcriptions
Wizard of Oz Interface

• Enhanced ARPABET symbols to represent phenomena
  peculiar to non-native and children's speech
• Dental stops
• Unaspirated voiceless stops
• Negative VOT (prevoiced) stops
• Lispy /s/
• Glottalized /t/
• Long frication of /f/
• Trill
• Syllabic Consonants
• With a convention to represent vowel space with
  respect to English vowels
• Non-native vowels defined by the two nearest
  English vowels, with the highest vowel first
  (e.g., /iyih/).
• 82 phone-level transcriber agreement.
• Transcribers started with 100 overlap (each
  file transcribed twice), 25 after agreement was
  established.
• Sentences are transcribed by word-level
  alignments, with phone-level detail if there was
  pronunciation variation.

techie
tester

• Language Background Effects
• Difficulty associating words with pictures.
• Sometimes reading sentences was performed
  better than individual words by children who
  could read in Spanish but not English.
• Sounding out words with Spanish letter-to-sound
  rules.

• Corpus Statistics
• 256 Children recorded.
• 30,000 utterances.
• 40 hours of speech.
• 13 GB of speech data sampled at 44.1 kHz.

child

• Age/Grade Effects
• Position in school year is important (children
  learn a lot between the beginning and end of the
  school year).
• Younger children are more timid.
• Less social and reading experience.
• Less exposure to computers.

• Reading Tactics
• Sounding out words generally helped children.
• Mispronunciations when a subword portion is
  confused with another word (e.g., once, using).
• Confusion with the different sounds an
  orthographic symbol may have (e.g., now
  pronounced as no).

• Our recording setup was similar to our target
  application
• Two visible operators
• The techie controls the presentation of stimuli
  and monitors recording quality.
• The tester instructs and guides the child.
• The target application would include just the
  child and computer.
• 20 min. per child maximum.
• A secure database stores child demographic info
  and speech recordings
• Age, grade, English level,
• native language, language used at home, language
  used with friends,
• parents' native language, parents' birthplaces.
• Accommodations for children.
• 15 children recorded per day (1.9 hours of
  recorded speech per day).
• Stimuli pictures, colors, alphabet, numbers,
  words, and sentences.

Acknowledgements

• Higher Level Phenomena
• Using a, an, some in picture naming.
• Perhaps due to grammatical differences in English
  and Spanish.
• Verb tense changes when reading sentences.
• Formation of contractions from long forms (but
  not vice-versa).
• Reanalysis of sentence after the child realizes
  a mistake he/she has made.

• Pronunciation Variation
• Read speech is slower.
• Long breaks in fricatives followed by stops
  (e.g., s-tart).
• Long liquids, nasals, and fricatives.
• Syllables spoken slowly (e.g., a-long).
• Final consonants may be delayed or dropped
  (e.g., par-t or par- ).
• Difficulty with am and an in isolation.
• At times, children speak in an exaggerated
  voice.

This project is supported in part by the NSF. In
addition, this work would not be possible with
out the hard work of transcribers Daylen Riggs
and Nathan Go the patience and bilingualism of
Kimberly Reynolds and Blanca Martinez the
careful recordings of Erdem Unal, Vivek
Rangarajan, Shiva Sundaram, Yirong Yang, Jinjin
Ye, and Yijian Bai and the planning of Larry
Casey and Christy Boscardin.