mohsen-goodarzi

> https://github.com/berndie/tensorflow_wheels/releases/download/v1.15.2/tensorflow-1.15.2-cp36-cp36m-linux_x86_64.whl Thank you so much! It worked great in a docker container in a virtual server!

I'm having exactly the same problem (not pronouncing the final word of sentences) in Persian. I've used a 14h dataset.

> In all my datasets there is no beginning/ending silence, ... Is it a must?

@csukuangfj I noticed that the `accent` column in `cv-corpus-5.1-2020-06-22` contains just single word values but in `cv-corpus-9.0-2022-04-27-nl`, they are two words separated by space like: `Nederlands Nederlands` or `Frans Nederlands`....

Of course not. Please see PR #743.

Why these results are still far behind conformer-s with 10M params from [conformer paper](https://arxiv.org/pdf/2005.08100.pdf) (2.7/6.3)?

One more thing: Here #574, it is shown that shuffling training cuts in prepare.sh is beneficial for reducing WER, and smother training loss curves, but isn't shuffling already enabled via...