bofeng huang

Hi @anautsch, thank you very much for your explanations ! And yeah, my question was exactly why using the arbitrary lognorm but not one with the parameters fitted on dataset....

@anautsch nice! Look forward to our discussion :)

Hi @anautsch , As discussed last week, I came back with the tests on more datasets : - MiniLibriSpeech - LibriSpeech - CommonVoice IT - CommonVoice FR I used `batch_size=32`...

@anautsch thanks for the rewriting. I didn't realize the dynamic sampler with beta distribution has the less initialization time (except on LibriSpeech). Just when fitting the beta distribution on datasets,...

Yeah it was the same "total time" used in the tutorial. > Interesting that fewest padding is not always fewest batches! I think this is related to datasets, #buckets and...

There are also some other changes including: - Change `_ex_lengths` from dict to np array - Isolate the function `_get_bucket_ids` which distributes the examples into buckets by bucket boundaries, to...

Hi @anautsch , I've run some real tests on LibriSpeech-ASR-CTC-wav2vec recipe with different setups on a single rtx8000. `num_workers=1` is added to the defaut recipe to get a little speed...

So I modified to fill up the batches by `batch_size` in order to get the same number of batches and ended up with the following results. | | test clean...

Hi again @kahne , `soundfile.write` accepts an array of `(T x C)`, but `_wavform` here has `(C x T)`. Fixed here in 2nd commit

Hi, `is_audio` arg has been added to `get_zip_manifest` for audio/fbanck in the 3rd commit