Research / Language technology for under-resourced languages / [11]
Galician speech recognition in 2022: two systems, and too little data
Galician's Language Technologies in the Digital Age. IberSPEECH 2022, 2022.
Only two speech recognisers supported Galician in 2022. On a small set of Galician TV programmes, the University of Vigo system produced better transcripts than Google Cloud Speech-to-Text in every case, and resources for the language remained scarce.
The question
Galician has about 1.9 million speakers and is co-official in Spain, yet it appears on fewer than 0.1 % of websites. This study, done within the European Language Equality project, asks where speech recognition for Galician stood in 2022.
What we did
We surveyed the resources available for the language and compared the two recognisers we found with Galician support: the one developed at the University of Vigo and Google Cloud Speech-to-Text. Both transcribed programmes from four Galician production companies, and the transcripts were scored by word error rate.
What we found
Data and resources for Galician grew markedly in the four years before the study, but they remain limited in scope: 92 % of the corpora and resources indexed are text only, and audio is rare.
The University of Vigo system produced better transcripts than Google’s on every programme. Substitutions were the most common error in both. A word error rate below 20 %, the level Spanish public television accepts for automatic subtitles, is reachable when the audio is controlled and has few speakers.
Why it matters
Speech technology now depends on large amounts of data. Without a substantial effort to create resources, above all audio, Galician risks being left behind. The fuller picture is in the Language Report for Galician.
Cite
Ramírez Sánchez J.M., Docío-Fernández L., García-Mateo C. Galician's Language Technologies in the Digital Age. IberSPEECH 2022, 2022.