joazoa
joazoa
@lin-nerd did you get this to work in the end ?
@b-flo can you share the config you used ? (I suppose it's for the new standalone version?)
Thanks for the super quick reply! For librispeech if possible, but librispeech-100 would help as well. (any other config for the new version would help me as well)
Thanks! I will give it a try.
I have been running this for a week, seems to work fine for me. Transformer decoding speed ~x4, training speed on small 27m conformer about 5% faster.
In the mean time i found an issue, LER comparison will include UNK tokens that were added for padding which will lead to unexpected results where the WER is 0...
Can you explain a bit more what you mean ? You can add more phrases to your lexicon before decoding and you could boost their them by using a language...
I think you can, but i don't think it will be fast enough. https://github.com/danpovey/pocolm/blob/master/docs/motivation.md https://github.com/kpu/kenlm/issues/98 SRILM also has something for it.
You can find the model here and some alternatives. https://github.com/flashlight/wav2letter/tree/main/recipes/rasr You can modify the file, but you will need to retrain from scratch. They did not provide the training configuration...
I am experiencing NAN when using mixedprecision only. The same datasets work well with all other architectures and |I do not not have NAN problems with mixedprecision disabled. I was...