Start with the job log. It records what Whisper did and usually names the problem. See The Job Log for where to find it.

The V100 GPUs may be busy. Open the Job Details window and read Reason, which says what the job is waiting for, then check the history again later. Do not submit the same job a second time; see Submit and Monitor the Job.

The job's status is Timeout and the output may be incomplete, so check the files before reusing them. Run it again with a longer maximum runtime, up to the form's limit of 7 hours and 59 minutes, or process fewer files at a time.

Select the spoken language explicitly instead of leaving the field empty, and run the recording again. For multilingual recordings, split the material by language when possible.

Run the job again with Skip alignment selected. The transcript is still produced, but word-level and subtitle timings may be less precise.

Confirm that Run speaker diarization was selected. Diarization is an estimate and can struggle with overlapping speech, noise, or similar voices, and speaker numbers must be matched to real names by hand.

If you chose the pyannote 3.1 model, confirm that you supplied a valid Hugging Face token. NB: This is a legacy model which will be removed in the near future.

Check in order:

1. that the job has finished successfully, with status Completed in the job history;
2. that you are looking in the output directory you selected on the form; and
3. the job log, for an error.

Two addresses, depending on what you need:

When you write to hpc@rug.nl, include:

• the Job ID from the job history;
• the job's status;
• the task, language, and maximum runtime you selected;
• the type and rough duration of the recording; and
• a short excerpt of the error from the job log.

Do not send recordings, transcripts, or your Hugging Face token. If an excerpt from the log contains personal data, remove it before sending.

What is the huggingface folder in my /scratch/$USER directory?

It holds models the app downloads from Hugging Face: the models that align the transcript with the audio word by word, and, if you choose the pyannote 3.1 diarization model, the models it needs. The app keeps them in scratch rather than your home directory, because they would quickly fill your home quota.

The folder contains only model files, never your recordings or results. It counts towards your scratch quota. You can delete it at any time; the next job downloads the models it needs again, so it will take a little longer to start.