Abstract / topic summary
As Germany’s central film archive, the Federal Archives of Germany stores over 250.000 documentaries and feature films that are continuously being digitised and made available online. The depth of archival descriptions provided for these films, however, varies greatly. While some films have been described in great depth, most films have so far received little archival meta data such as content, mentioned persons, places, or institutions. This lack of searchable data poses an obstacle for research as films pertaining to specific keywords cannot be found. Thus, automatically transcribing the spoken text makes the contents of a film searchable without having to manually transcribe or summarize it.
To provide these transcriptions, we leverage novel machine learning techniques to automatically turn speech into text. This so-called Automatic Speech Recognition (ASR) enables the processing of films in significantly shorter time frames and makes the archive material more accessible. Using WhisperX, an open-source speech recognition model, we are able to automatically deliver good-quality transcriptions. We have demonstrated feasibility in an initial proof-of-concept stage, before developing a subsequent version for product development. Our results are evaluated using common state-of-the-art metrics in the field, and we further explore the possibility of predicting transcription quality through the models’ word probabilities. Additionally, our pipeline allows for configuration of different pre- and post-processing methods which were investigated in the scope of our project.
Implementing and connecting these components with our existing systems and infrastructure, we can provide readable transcriptions of films at larger scale and, thus, make the content of the audiovisual archive material more accessible for the users of the Federal Archives.
Speaker summary
Vy Nguyen
Vy obtained degrees in Media Informatics and Data Science. She gained professional experience as a software developer in digital commerce and worked on different projects in the field of data analysis and data science. In her last position, she was a researcher in the field of AI and machine learning before joining the German Federal Archives in Berlin as a data scientist in 2025.
