MXNet implementation of RNN Transducer (Graves 2012): Sequence Transduction with Recurrent Neural Networks
Compile RNNT Loss Follow the instructions in here to compile MXNET with RNNT loss.
Extract feature
link kaldi timit example dirs (local
steps
utils
)
excute run.sh
to extract 40 dim fbank feature
run feature_transform.sh
to get 123 dim feature as described in Graves2013
Train RNNT model:
python train.py --lr 1e-3 --bi --dropout .5 --out exp/rnnt_bi_lr1e-3 --schedule
Default only for RNNT
python eval.py <path to best model parameters> --bi
python eval.py <path to best model parameters> --bi --beam <beam size>
CTC
Decode | PER |
---|---|
greedy | 20.36 |
beam 100 | 20.03 |
Transducer
Decode | PER |
---|---|
greedy | 20.74 |
beam 40 | 19.84 |