Automatic Speech Recognition of Finnish-Swedish Dialects: A Comparison of Three Cutting-Edge Technologies

Leonardo Espinosa-Leal*, Kristoffer Kuvaja Adolfsson, Andrey Shcherbakov

*Corresponding author for this work

Research output: Chapter in Book/Report/Conference proceedingConference article in proceedingsScientificpeer-review

Abstract

This paper explores the performance of two different automatic speech recognition models for the Finnish-Swedish language. The first model, Whisper V1 released by OpenAI and the second, the KBLab model trained using a large dataset by the National Library of Sweden. These models were trained initially using data from the Swedish language from Sweden, and the results were compared with previous work trained using a dataset of Finnish-Swedish audio. Our results indicate that general models perform at the same level, opening up the possibility of using these in Finland for the inclusion of the Finnish-Swedish minority.
Original languageEnglish
Title of host publicationSmart Technologies for a Sustainable Future
Subtitle of host publicationProceedings of the 21st International Conference on Smart Technologies & Education
EditorsMichael E. Auer, Reinhard Langmann, Dominik May, Kim Roos
Place of PublicationCham
PublisherSpringer
Pages309-315
Number of pages7
Volume2
ISBN (Electronic)978-3-031-61905-2
ISBN (Print)978-3-031-61904-5
DOIs
Publication statusPublished - 2024
MoE publication typeA4 Article in a conference publication
Event21st International Conference on Smart Technologies & Education (STE-2024) - Helsinki, Finland
Duration: 6 Mar 20248 Mar 2024

Publication series

SeriesLecture Notes in Networks and Systems
Volume1028
ISSN2367-3370

Conference

Conference21st International Conference on Smart Technologies & Education (STE-2024)
Country/TerritoryFinland
CityHelsinki
Period6/03/248/03/24

Fingerprint

Dive into the research topics of 'Automatic Speech Recognition of Finnish-Swedish Dialects: A Comparison of Three Cutting-Edge Technologies'. Together they form a unique fingerprint.

Cite this