I reproduced the approximate librispeech TER from the original paper in around 40 epochs, it's the 1.6GB epoch 40 model here: https://talonvoice.com/research/
If you're training new English models these days, you should definitely try streaming convnets - it's a much better architecture. Faster, smaller, more accurate, and better supports online streaming. Some of my results with streaming convnets on English are here: https://github.com/facebookresearch/wav2letter/issues/577
I reproduced the approximate librispeech TER from the original paper in around 40 epochs, it's the 1.6GB epoch 40 model here: https://talonvoice.com/research/
If you're training new English models these days, you should definitely try streaming convnets - it's a much better architecture. Faster, smaller, more accurate, and better supports online streaming. Some of my results with streaming convnets on English are here: https://github.com/facebookresearch/wav2letter/issues/577