Chain alignments used for training

[Neuroxic]

I am trying to use the alignments generated by steps/nnet3/align.sh from my first chain model to train my next chain model (with the same data and model configuration). I want to investigate if better alignments have a noticeable impact on the WER.

Unfortunately, when creating a new tree, I get the error message "Alignments have wrong size X vs. Y".

Have I overlooked something when creating the alignments or can alignments of chain/nnet3 models not be used for training?

Many thanks!

[danpovey]

You may be able to fix this by setting --chain.alignment-subsampling-factor=1 to steps/nnet3/chain/train.py
it default to 3 which is necessary because GMM alignments are at 100 frames per second, not 33