Hlavní stránka Info portálu ZČU

Scientists from the Faculty of Applied Sciences at the University of West Bohemia in Pilsen (FAV UWB) are developing systems capable of automatically transcribing and searching the content of extensive audio archives. Their technologies make the work of historians and the general public easier, enabling access to thousands of oral history recordings, including Holocaust testimonies.

The speech recognition systems developed by the team from the Department of Cybernetics and the NTIS Research Center have gradually improved from an initial error rate of 40% to around 10%. In addition to more accurate transcription, the researchers have developed a search tool that can locate words in various grammatical forms, phonetically similar expressions, and semantically related passages. “In the latest version, semantic search can locate segments relevant to a query even when the specific word does not appear in the recording,” explained Jan Švec.

The research team has also developed an algorithm for automatically generating questions discussed in the recordings. These are displayed in a clear format within the user interface, providing direct access to relevant parts of the interviews. According to Pavel Ircing, the technology can be applied not only in making historical archives more accessible but also in areas such as podcasts, news reporting, and working with scanned documents.

The development of advanced technologies began with the participation of Pilsen researchers in the MALACH project (Multilingual Access to Large Spoken Archives) in 2005. Within this international collaboration, Czech and foreign teams sought to develop effective methods of speech recognition and information retrieval. The aim was to create a tool that would help both professional and non-expert users navigate the vast collection of recordings gathered by fieldworkers of the foundation established by Steven Spielberg after the release of his famous film Schindler’s List.

At that time, the FAS experts focused mainly on systems for automatically transcribing recordings in Slavic languages and Hungarian. “When the project ended, we had not yet achieved fully satisfactory results. We were dealing with 115,000 hours of recordings in 32 languages, about half of them in English. To make matters more complicated, for the vast majority of speakers, English was not their native language, so we had to cope with different pronunciations and accents,” said Pavel Ircing.

For this long-term work, Pavel Ircing and Jan Švec received the Steven Krauwer Award on Tuesday, September 30, in Vienna, during the annual conference of the European CLARIN infrastructure. “Although the two of us received the award, the results are thanks to a large team of colleagues from the Department of Cybernetics and the NTIS research center. Special thanks go to Josef Psutka, Martin Bulín, and Jan Leheček,” emphasized Pavel Ircing.