Automatic speech recognition: system variability within a sociolinguistically homogeneous group of speakers
Year: 2023
Type: conference-paper
Abstract: Orthographic transcription is a time-consuming task that is increasingly being automated for various applications.A growing body of work on algorithmic bias has highlighted variability in automatic speech recognition performance across demographic groups, but little research has focused on variability within a sociolinguistically homogenous group.This paper considers this issue and explores the extent to which phonetic properties of the voice (e.g.f0, formants) can predict how well an automatic system performs.Recordings of 99 young, male Southern British English speakers were automatically transcribed using Amazon Transcribe, and word error rate was calculated.A wide range of error rates (between 11-33%) is observed.However, of all acoustic, temporal and voice quality measures analysed, only articulation rate has a significant impact on error rate.System performance therefore remains difficult to predict within demographic groups.
Authors: Harrington, Lauren, Hughes, Vincent
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
SDG: Reduced inequalities
Open Access status: green
OpenAlex ID: W6980108438