Piotr Żelasko
Piotr Żelasko
Also some examples of how the batches look like in our custom setup can be seen here.. https://github.com/lhotse-speech/lhotse/pull/234
Thanks @ejguan! This looks pretty cool! I think I understand the whole flow. The GroupBy operator and matching by batch index is a cool idea -- but how does GroupBy...
Thanks! It’s good to know. I think time-domain is fine, it seems on-the-fly audio reading and feature extraction is efficient enough in the setups we tried so far.
You might want to try decreasing num_jobs and max_duration in that script. I don't know why this sometimes happens with `compute_and_store_features_batch` but I've seen that before.
... there's also the question of storing/restoring the state to resume training from a checkpoint.
It looks like a bug, what you're suggesting seems correct. Do you mind making a PR with your fix and a unit test to check that it's doing the right...
Thanks for letting us know! Actually I think I can add a check before lilcom compression so that when it detects only (or mostly) non-negative values/narrow value range, it will...
Hmmm I'm not sure we want to do that by default. We could have a flag to turn off SSL for this recipe, e.g. using this solution https://stackoverflow.com/questions/33770129/how-do-i-disable-the-ssl-check-in-python-3-x Would you...
@oshindow can you try the solution above?
Hmm we had a similar issue here, but it's already fixed https://github.com/lhotse-speech/lhotse/issues/753 Can you print the contents of the sampler's `state_dict` that's saved on disk? It would also be helpful...