You have to get the "prons.n.gz" from kaldi/get_prons.sh and concate the files together and named as all_prons, also gets the relative features as cmvned_feats.ark
Then you can use the scripts inside the directory like
./script/train_origin.sh ./feat 100 ./model ./log 1 80000
Since the criteria of the script is
./script/train_origin.sh feat-dir memory-dim model-dir log-dir gpu-num max-training steps
You may change the model structure by modifying the function inference
, loss
in src/audio2vec\_train.py
, if you want to modify the cell or the
elemental connection of sequence-to-sequence structure, please visit src/seq2seq.py
Starts training by using script/train\_origin.sh
Evaluation by using src/audio2vec_eval.py