Early results from survey exploring transcription processes
The Transcription Acceleration Project (TAP) is a CoEDL cross-disciplinary project aimed at identifying the workflow of linguists during transcription, and developing assistive tools to accelerate that process.
Led by Janet Wiles and Nick Evans, the project involves technology experts based at the University of Queensland and linguists across all CoEDL’s four university nodes.
Postdoc Gautier Durantin conducted a survey of linguists earlier this year to get a picture of linguists’ practices during transcription, especially how they use technology during that process.
There were 51 respondents to the survey and early results show that the average time spent transcribing per year is 152 hours, with the average time to transcribe taking 39 minutes for one minute of audio.
Dr Durantin says this means that on average the linguists who responded to the survey are transcribing only a fraction of their corpus each year.
“It emphasizes how important it is to accelerate the process, given that training language models or extracting linguistic information automatically usually requires hundreds of hours of transcribed data,” he says.
The survey has also revealed some important tricks and techniques that linguists use to speed up their transcription process. Further results of the survey will be published later this year.
CoEDL Chief Investigator Professor Janet Wiles, says despite their efforts, linguists are losing the race to document the 3000–4000 languages that are likely to disappear this century, so it is important everything is done to support the documentation process, including improvements to existing transcription software and the development of improved technology.
She says while the problem from the perspective of linguistics is the time it takes to transcribe recordings, commonly referred to as the “transcription bottleneck”, the critical problem for computer scientists is, “how do we get a computer to listen like a human?”
She believes immediate gains might be found in improving the interface between linguist and software. “Greater cooperation from technology experts and linguists is likely to pay big dividends in this context,” she says.
Top image: Kelsey Daniels working with Tun Lynn Yanam to transcribe Soq.