Uncovering bias in ASR systems:Evaluating the performance of Wav2vec2 and Whisper for Dutch speakers

The study evaluated two speech recognition systems, Wav2vec2 and Whisper, for potential biases for Dutch speakers.Results obtained by evaluating on the JASMIN corpus revealed biases against non-native speakers, children, and the elderly,with (slightly) better performance for women. The study emphasizes the need for ASR systems to handle variations in speakingin order to reach equal performance among all users.

Verfasser: Fuckner, Marcio
Horsman, Sophie
Janssen, Iskaj
Wiggers, Pascal
Dokumenttyp: conferenceObject
Erscheinungsdatum: 2024
Sprache: Englisch
Permalink: https://search.fid-benelux.de/Record/base-26633875
Datenquelle: BASE; Originalkatalog
Powered By: BASE
Link(s) : https://research.hva.nl/en/publications/21351931-34f3-4715-95c6-b24c0f699505