Titel

The solution needed to be powerful enough to run modern AI models, operate entirely on local infrastructure, and give the university full control over processes and data. These core requirements were quickly met. However, as discussions with users progressed, expectations grew step by step: speakers had to be identifiable, timestamps inserted, and translations into other languages enabled. What began as a small assignment evolved into a sophisticated project combining advanced features, computational efficiency and digital sovereignty.

Today, the tool follows a clearly defined workflow spanning a Linux workstation and the local HPC system LEO5. After an audio file is transferred from a workstation at the Office for Open Science within the Faculty of Social and Political Sciences to the university’s supercomputer, the audio is analysed and segmented by speaker. The tool then automatically transcribes each segment. If required, the completed transcript can subsequently be translated into other languages.

Screenshot of the new AI transcription tool. In the left column: all processes running on the CPUs, including the system’s total CPU utilisation, which is very low here. Beneath the green bar on the far right: the task currently being computed — in this case, the transcription itself. In the right column: the utilisation of GPU 1 in terms of compute capacity (99% utilisation) and memory usage (23,293 MiB of 24,576 MiB; MiB = mebibyte) during the actual transcription process.

The AI solution deliberately builds on established, well-documented components, which Andreas Lindner specifically adapted for use on LEO5:

• Transcription is performed using a Whisper model: https://huggingface.co/docs/transformers/en/model_doc/whisper
• Speaker diarisation – the process of assigning speech segments and timestamps to individual speakers – is handled by models from pyannote: https://huggingface.co/pyannote
• Translation of transcripts is carried out using a Tower model developed by Unbabel: https://huggingface.co/collections/Unbabel/tower
• Data encryption and decryption are implemented using gocryptfs, ensuring secure handling of sensitive information: https://www.baeldung.com/linux/gocryptfs-encrypt-decrypt-dirs
• To guarantee a clearly separated and reproducible runtime environment on LEO5, the speech-to-text workflow runs inside a software container.

Left column: speaker diarisation in progress. Right column: utilisation of GPU 1 — memory usage (2,709 of 24,576 MiB used). The compute utilisation of GPU 1 is at 70%.

The solution is now technically operational and fully documented. It integrates into the existing research infrastructure, relieves researchers of the time-intensive transcription process, and safeguards data sovereignty. AI Factory Austria (AI:AT) uses the tool internally to transcribe video conferences. Incidentally, the interview with Andreas Lindner on which this article is based was itself transcribed using the new tool on LEO5.

An additional benefit of the system is that transcription tasks are not time-critical. They are automatically scheduled into computational gaps between other projects. This improves utilisation of the HPC systems, which consume power even when idle. By putting otherwise unused capacity to productive use, overall energy efficiency increases.

With this new tool, the Faculty of Social and Political Sciences now has access to an AI-supported solution that simplifies research workflows, preserves data sovereignty, and demonstrates how modern technology can be deployed responsibly.

Since automated transcription is needed not only in Innsbruck but in many other institutions, Andreas Lindner has made the corresponding GitLab repository publicly available. It may therefore also be useful for other operators of high-performance computing clusters who wish to implement a similar setup:

https://researchdata.uibk.ac.at/records/z877w-c6110
https://doi.org/10.24433/CO.0416787.v1

STT workflow manager: the user’s initial steps when running the script, for example selecting the language and the number of speakers (at the bottom).