Luca Della Libera

icon indicating a website link https://www.linkedin.com/in/luca-della-libera

icon company Concordia University icon location Montréal, Québec, Canada icon location PhD student at Concordia University and Mila, currently working on speech processing.

Results 4 comments of Luca Della Libera

[Bug]: AttributeError: Can't pickle local object 'dataio_prep.<locals>.audio_pipeline'

Can you try to add these lines at the beginning of your training script: ```python import multiprocessing multiprocessing.set_start_method("fork") ```

[Bug]: Special token IDs (BOS, EOS, etc.) not matching the tokenizer's configuration

I'm also not sure, it might lead to other incompatibilities. I think we should keep this in mind for future recipes and be more careful about setting the correct token...

DDP at eval stage + DDP metrics

The small variations in CER and WER might be related to the fact that when the dataset is not divisible by the number of processes, a few samples are duplicated...

[WIP] PyTorch K-means for discrete audio tokens extraction

![tsne-ljspeech-batch](https://github.com/speechbrain/speechbrain/assets/34525085/cfa8c00f-f5b7-4213-b338-72151a5f1931)