SpeechbrainEmbeddings fails when run in dask for the first time

When running speechbrain-ecapa-voxceleb for the first time, it fails if on dask.

Multiple workers are calling load_model at the same time and as the model files do not exist (first time running the pipeline), speechbrain tries to download them from huggingface. It uses a cache in ~/.cache/huggingface. Error rise as multiple worker access simultaneousely to the same files.

Workaround: The first time you run speechbrain-ecapa-voxceleb, do it without the -l ... option, at least until the files are downloaded and the computations starts. You can then stop the execution and run with the -l option.