Aarhus University Seal

Coder bias

Paving the Way for a More Reliable Linguistics

Speech transcriptions, the decoding of sounds in letters or other visual symbols, are important means for investigating pronunciation. For instance, sounds of a particular dialect can only be statistically analysed if they have been transcribed. Transcriptions are also used in foreign language education, speech therapy, and language counselling (like in asylum requests). Transcriptions are currently often made by speech processing software, but have to be corrected manually: it cannot fully replace human codings. Some sounds cannot be analysed by speech processing software at all. But are human codings reliable? The answer is probably no. Recently, it turned out that linguistically subjects are easily biased towards an overall accent they perceive in a speaker, such that coders exposed to a certain accent may expect and therefore actually hear prototypical characteristics of this accent—even if these are not present in the phonetic signal. Even professional coders turned out to be subject to coder bias. The scope of this deviation in formal transcription tasks is still unknown, neither has a standard procedure been developed in order to arrive at better results. Since human codings should be as accurate and as reliable as  possible, this research aims at the development of a “gold standard” for linguistic transcriptions.

As part of the project there will be a conference i September 2014 "Bias in Auditory perception".

Project leader: Marjoleine Sloos (NL) msloos@cas.au.dk