Jerry Schnepp, Ph.D
Chair, Computer Science and Cyber Security
Sign Language Synthesis
Part of a larger project to improve Deaf/Hearing communication through 3D animation.
Animations must not only be understandable but also be acceptable to the Deaf community.
2
Why we need it
English is a second language for most members of the deaf community
The average reading level of Deaf adults is between the third and fourth grade level
4
Face-to-face testing challenges
Language barrier, Scheduling, Cost, Recruitment, Locality
7
Remote testing with SignQUOTE
Interface entirely in Sign, Test at any time, Simultaneous sessions, Reduces cost
8
IRB
A love/hate relationship
9
Description
10
Synthesizing American Sign Language (ASL) via computer animation is an essential component of any automatic English-to-ASL translation system. The goal of this study is to evaluate the legibility, grammaticality and acceptability of synthesized ASL. Without legibility, the animations are incomprehensible. Without grammaticality, the intended meaning may be garbled and incorrectly perceived by an observer. Without acceptability, Deaf users will avoid using any system that produces these animations. Feedback from study participants will allow the investigators to review and revise underlying motion models for improving the synthesis process.
11
In the instructions (Appendix D) the participants are informed that several of the questions involve recording their responses via their computer's webcam. Recordings for each participant are stored on a secure server. The PI never sees the recordings. Instead a certified ASL/English interpreter views the recordings and voices the responses that are transcribed by the PI. Once the transcription is complete and checked for accuracy, the recordings are destroyed by deleting the files.
The certified interpreter is to be determined. When the PI hires the interpreter, the PI will only select an interpreter that has been certified by the Registry of Interpreters of the Deaf (RID). For the high standards of confidentiality required of certified interpreters, see http://rid.org
Only the certified interpreter will see a participant’s face. Even if the interpreter recognizes the participant, which is unlikely, since the participants are located throughout the country, the code of conduct of certified interpreters forbids them from identifying the participant by name. The interpreter will be voicing the open-ended responses to the researchers for transcription, but the researchers will never see the participants’ faces. Note that this is a higher standard of confidentiality than can be maintained in face-to-face usability tests.
12
13
Method
14
Comparing F2F and Remote
The Mann-Whitney metric attempts to measure the likelihood that a difference in distribution between two sample populations arises from random variation.
A very low Mann-Whitney score (<.05) indicates that the differences in the two medians are statistically significant.
The scores do not indicate that these differences are significant, thus it is quite possible that the difference between the face-to-face and remote medians resulted from randomness in the samples.
15
Comparing F2F and Remote
The percentage of participants who gave substantive suggestions for improvement. Responses such as “It’s fine,” or “no comment” were omitted while responses such as “The brows should be up longer” or “She’s signing too slow” were deemed substantive suggestions.
The differences between the face-to-face and remote testing methods were not statistically significant, but the percentages were consistently higher in the remote scenario.
16
Jerry Schnepp, Ph.D
Chair, Computer Science and Cyber Security