Overview
- The joint project by LMU’s Center for Information and Language Processing and Bayerischer Rundfunk evaluated BR’s Betthupferl bedtime stories recorded across Bavaria’s major dialect groups.
- Three speech-recognition model families were tested, including OpenAI’s Whisper, to compare performance on dialect recordings with Standard German samples.
- Error rates on dialect speech were markedly higher than on Standard German, with BR reporting that sentence meaning was often lost in transcription.
- Whisper handled Swiss German relatively well in the tests, yet it did not resolve the significant errors on Bavarian dialects.
- The validated findings were presented this week at the Interspeech conference in Rotterdam, and the teams propose expanding dialectal datasets to improve accuracy and support automated subtitling of dialect broadcasts.