Datasets:
include bias from voting instead of transcription
#6
by
jsalsman
- opened
Describe the inherent bias against authentic intelligibility of accented speech for speech-to-text and pronunciation assessment applications when "thumbs" upvotes and downvotes are used instead of transcriptions typed by listeners. Included are the immigration issues that "ruined thousands of lives" according to the BBC, Pearson's correction of the problem for their Versant high-stakes assessment, a discussion from the academic pronunciation teaching literature, and a link to Mozilla's EM Lewis-Jong discussing the trade-offs.
Minor typo: in "Mozilla's EM Lewis-Jong discussed the trade-of of..." should be "the trade-off of...."