Automated Analysis of Speech Data from HALA: A Voice Onset Detection Tool for Assessing Language Attrition
4
조회
0
다운로드
0
인용
0
참고문헌
키워드
초록
This study presents the development of an automated data analysis tool for the Hawai‘i Assessment of Language Access (HALA), a psycholinguistic task designed to assess lexical access speed as an indicator of language attrition. While HALA has been instrumental in studies of language attrition, its manual response time (RT) measurement process is labor-intensive and limits the scale of studies. To address this issue, we developed an automated tool that determines the time interval between the beep that accompanies the visual stimulus and the onset of a respondent’s verbal response, leveraging pitch patterns. Validation against 3,755 human-coded trials from 146 participants showed that 32.68% of automated RTs deviated by less than 100 ms, 33.9% by 100–500 ms, and the remainder by more than 500 ms. Qualitative analysis revealed that deviations often occurred when participants produced filled pauses such as uh or oh prior to naming the target word. These results suggest that while automation is feasible, further refinement is needed to better align with human measurement. Based on these findings, we discuss the effective use of this tool as a supplement to, rather than a replacement for, human annotation.
Abstract
This study presents the development of an automated data analysis tool for the Hawai‘i Assessment of Language Access (HALA), a psycholinguistic task designed to assess lexical access speed as an indicator of language attrition. While HALA has been instrumental in studies of language attrition, its manual response time (RT) measurement process is labor-intensive and limits the scale of studies. To address this issue, we developed an automated tool that determines the time interval between the beep that accompanies the visual stimulus and the onset of a respondent’s verbal response, leveraging pitch patterns. Validation against 3,755 human-coded trials from 146 participants showed that 32.68% of automated RTs deviated by less than 100 ms, 33.9% by 100–500 ms, and the remainder by more than 500 ms. Qualitative analysis revealed that deviations often occurred when participants produced filled pauses such as uh or oh prior to naming the target word. These results suggest that while automation is feasible, further refinement is needed to better align with human measurement. Based on these findings, we discuss the effective use of this tool as a supplement to, rather than a replacement for, human annotation.