End-to-End Speech Recognition using RNN-Transducer

File description

• eval.py: rnnt joint model decode
• model.py: rnnt model, which contains acoustic / phoneme model
• model2012.py: rnnt model refer to Graves2012
• seq2seq/*: seq2seq with attention
• rnnt_np.py: rnnt loss function implementation on mxnet, support for both symbol and gluon refer to PyTorch implementation
• DataLoader.py: data process
• train.py: rnnt training script, can be initialized from CTC and PM model
• train_ctc.py: ctc training script
• train_att.py: attention training script

Directory description

• conf: kaldi feature extraction config

Reference Paper

• RNN Transducer (Graves 2012): Sequence Transduction with Recurrent Neural Networks
• RNNT joint (Graves 2013): Speech Recognition with Deep Recurrent Neural Networks
• E2E criterion comparison (Baidu 2017): Exploring Neural Transducers for End-to-End Speech Recognition
• Seq2Seq-Attention: Attention-Based Models for Speech Recognition

Run

• 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

Evaluation

Default only for RNNT

• Greedy decoding:

  python eval.py <path to best model parameters> --bi

• Beam search:

  python eval.py <path to best model parameters> --bi --beam <beam size>

Results

• CTC

  Decode	PER
  greedy	20.36
  beam 100	20.03

• Transducer

  Decode	PER
  greedy	20.74
  beam 40	19.84

Requirements

• Python 3.6
• MxNet 1.1.0
• numpy 1.14

TODO

• beam serach accelaration
• Seq2Seq with attention