National Library of Sweden develops AI model using historical parliamentary speeches

The National Library of Sweden is developing an AI model, utilizing a newly arrived hard drive containing digitized recordings of historical parliamentary speeches. According to Leonora Vesterbacka, a Senior Data Scientist at the library, this serves as excellent training material for speech-to-text models.

The project aims to compile examples of various dialects, accents, and distinct speech patterns to create an AI model that is accessible and free for all, enabling the development of speech-to-text services in Swedish. Vesterbacka notes practical applications, such as dictating and transcribing medical records.

The library has partnered with the Institute for Language and Folklore in Uppsala, which houses tens of thousands of hours of archived speech, including recordings over 100 years old. Annette Torensjö, the institute’s archive manager, highlights the importance of catering to elderly individuals who may need assistance with phone services or hospital calls.

While there is a wealth of thick dialects in the archives, material featuring non-Swedish speakers is lacking. Vesterbacka stresses that a transcription service must cater to all residents of Sweden, not just those who speak standard Swedish.

Overall, the initiative aims to broaden the accessibility of AI speech technologies across the diverse linguistic landscape of Sweden.