Salme - Speech recognition in courts

This information was sourced from the European Commission, Joint Research Centre (2026): PSTW: Public Sector Tech Watch latest dataset of selected cases. [Dataset] doi: 10.2905/JRC.J1MWC9R PID: http://data.europa.eu/89h/e8e7bddd-8510-4936-9fa6-7e1b399cbd92. Licensed under Creative Commons Attribution 4.0 (CC BY 4.0).

Speech recognition software enables the transcription of a written transcript of a hearing ('the transcript'), which is used to draw up the minutes of the hearing. COMPRISE H2020 project is experimenting a solution for this. Tilde – a leading European language technology company – in collaboration with the Estonian national Centre of Registers and Information Systems (RIK) has developed a solution for automated transcription of court hearings in all 9 national and regional courts of the country. The solution has brought significant improvements to the efficiency and speed of producing court session recordings and protocolling.

Speech recognition models have been adapted for the rather rich domain content – more than 800 hours of transcribed audio files have been used for the development of the acoustic models and over 800 million words worth of textual data have been used for the custom language model development. As a result, the solution which provides for real-time, as well as offline speech recognition has high quality of speech recognition, where word error rates (WER) vary between 8 and 15%, and this is a very good result given the actual conditions at court hearings. This solution is built to run on the RIK’s infrastructure thus safeguarding the security and availability to authorized users only.

Budget: Not disclosed