Ravnur — Open Faroese Speech-Recognition Resources Built from 135 Hours of Dialect-Balanced Recordings

The Ravnur project was assembled in 2019 to build open resources for automatic speech recognition (ASR) in Faroese, a language with no ASR tools at the time. It was funded by public and private backers, including the Faroese government, and run by the Talutøkni Foundation with linguists, students from the Faculty of Faroese and IT staff.
Ravnur recorded 135 hours of speech from 433 speakers (249 women, 184 men) reading texts of various genres, covering all dialects and ages, with manual orthographic transcription and a pronunciation dictionary. BLARK 1.0 (Basic Language Resource Kit) was published on 1 July 2022, with a 109-hour extract, the Ravnursson corpus, curated at Reykjavík University and released in November 2022 under CC BY 4.0.
In the 2023 NoDaLiDa paper, models trained on the corpus reached 7.60% word error rate on the test set with a fine-tuned wav2vec2 XLSR-53 model (versus 14.04% for Kaldi LSTM). The authors also found that Whisper gave very poor results without fine-tuning on Faroese. Accessibility context: the only earlier Faroese speech synthesis (2005) is used by the visually impaired community, and the authors report strong demand for ASR among deaf, visually impaired and dyslexic people.
Limits: the sources report technical accuracy, not adoption. No user numbers or deployed accessibility product based on the resources were found, and the corpus consists of read speech rather than spontaneous speech.

Read the full analysis: https://mtd.setur.fo/en/2022/07/04/verkaetlanin-ravnur/