Automated generation of ‘good enough’ transcripts as a first step to transcription of audio-recorded data

Abstract

In the last decade, automated captioning services have appeared in mainstream technology use. Until now, the focus of these services have been on the technical aspects, supporting pupils with special educational needs and supporting teaching and learning of second language students. Only limited explorations have been attempted regarding its use for research purposes: transcription of audio recordings. This article presents a proof-of-concept exploration utilising three examples of automated transcription of audio recordings from different contexts; an interview, a public hearing and a classroom setting, and compares them against ‘manual’ transcription techniques in each case. It begins with an overview of literature on automated captioning and the use of voice recognition tools for the purposes of transcription. An account is provided of the specific processes and tools used for the generation of the automated captions followed by some basic processing of the captions to produce automated transcripts. Originality checking software was used to determine a percentage match between the automated transcript and a manual version as a basic measure of the potential usability of each of the automated transcripts. Some analysis of the more common and persistent mismatches observed between automated and manual transcripts is provided, revealing that the majority of mismatches would be easily identified and rectified in a review and edit of the automated transcript. Finally, some of the challenges and limitations of the approach are considered. These limitations notwithstanding, we conclude that this form of automated transcription provides ‘good enough’ transcription for first versions of transcripts. The time and cost advantages of this could be considerable, even for the production of summary or gisted transcripts.

Cite this work

No public copy is stored for this record, so only the metadata is cited.

Stored only when both papers are in the corpus — 63 citations exist upstream.

Papers that cite the same references — the relational fallback while embeddings are still backfilling.