Huahuan Zheng

Edited: with bug in my text processing fixed, this issue remains.

I believe this is not code-related issue. I've trained a 5-gram model with 47GB corpus, which takes around 3 hours, and it could be even faster since I set `-S...

@mthrok `apply_effects_tensor ` is OK, as I said in the bug description ``` # this won't produce error. torchaudio.sox_effects.apply_effects_tensor( torchaudio.load(file)[0], 48000, effects ) ``` So I think torchaudio should deprecate...

Hi @b-flo , I'm also working on this impl. I have implemented the section 3.1 with CUDA binding and make better perf. Do you have any idea about the section...

> I'm not sure about what you're asking? From what I understand, computation would be slightly slower with same settings but the memory reduction allows us to use larger batches,...

Is there a plan to support the packed layout logits of RNNT loss? Ref: Sec 3.1 https://arxiv.org/abs/1909.12415

Will do further investigation later :) As for the IO issue, I remember I have seen in somewhere that a thread block would instinctively load nearby memory whatever it is...

I've been following the fast_rnnt work for a while, but haven't make a successful **pruned** rnn-t training yet. They also have a paper about the implementation. https://arxiv.org/pdf/2206.13236.pdf

There is a similar loss function impl from K2 https://github.com/danpovey/fast_rnnt/blob/2c2dc4b96a6b9a8c0dbedada94cdee53a9337402/fast_rnnt/python/fast_rnnt/rnnt_loss.py#L159-L162 It seems they just add a small value to avoid log(0), which would also introduce errors in calculation. @pkufool @csukuangfj...

The only way I could figure out is to implement a custom function`logmmexp(a, b)`, where we need to compute `a_k + b_k` twice (Since we don't want to store the...