Groningen’s local Whisper service passes 1,000 university users

More than 1,000 University of Groningen staff and students are using a locally hosted Whisper transcription service, the university reported on 6 October. Most users come from its behavioural and social sciences, economics and business, and arts faculties. For a high-performance computing cluster more commonly associated with science, engineering and medicine, that is a revealing change in clientele.[1]

The service runs OpenAI’s speech-recognition model on Hábrók, the university’s computing cluster. Groningen says recordings remain within its infrastructure and are not sent to external AI providers. Researchers, students and support staff can use it without a service charge. The reported user total measures uptake; the announcement gives no total of audio hours processed or measured time saved.[1]

The attraction of this arrangement is easy to understand. An interview project needs computing power for a bounded task: turn a recording into text that can be searched, checked and analysed. The researcher does not necessarily need a permanent personal AI subscription, nor a workstation built around that occasional burst of work.

The service around the model

Groningen’s Digital Competence Centre began work on a local transcription solution in mid-2023. Its autumn 2025 launch material described requesting a computing account, following an installation guide and running a script. Jobs could be queued and their results collected later, while other work continued. That was already a different arrangement from sending an interview to an external transcription website.[2]

The current public guide describes a browser-based batch application with no installation or command-line work required. It also identifies the division of labour behind it: the Digital Competence Centre developed the concept, a data-science team built the early command-line version, and the high-performance computing team builds and maintains the portal application.[3]

That is an institutional product-development story as much as a model-deployment story. The model supplies a capability. Accounts, a usable interface, a queue and a support team make that capability available to people whose research expertise lies elsewhere. Each of those pieces has to fit an ordinary research task.

The default four-hour runtime is a ceiling, not a predicted completion time. A queued job may be waiting for a GPU. The guide cautions against submitting it twice: duplicate jobs consume extra resources and can overwrite the same output files. These are the practical consequences of sharing computing capacity.[4]

A computing service is also a place to leave

Hábrók’s scratch space is temporary and is not backed up. Groningen’s instructions therefore include downloading the needed results and clearing the recordings, transcripts and job log from the cluster.[6]

That final movement is part of the design. A researcher borrows computing capacity, brings back a working document and continues in the environment where the project’s material belongs. The cluster does not have to become the permanent home of the interview collection.

The university’s user figures show that this particular route has found an audience beyond its traditional computing disciplines. They do not isolate which feature attracted each person. But the documented service explains what those users are being offered: a familiar research job, supported locally, with the complex machinery largely behind a form.[1]

Report an error or suggest a correction