requirement: tensorflow >= 1.6
- preprocessing (./wsj_data)
we converted .wv1 files in WSJ corpus to .wav files using sph2pipe.
Features are extracted and saved as numpy file with preprocessing.py
- train (character-level model)
python train_wsj_sru_ifo.py // train i-SRU
python train_wsj_sru_conv_ifo.py // train i-SRU + conv
python train_wsj_lstm.py // train LSTM
- test (character-level model)
python test_wsj_sru_ifo.py // test i-SRU
python test_wsj_sru_conv_ifo.py // test i-SRU + conv
python test_wsj_lstm.py // test LSTM
- train (wordpiece model)
python train_wsj_sru_wordpiece.py // train i-SRU + conv model
python train_wsj_lstm_wordpiece.py // train LSTM model
- test (wordpiece model)
python test_wsj_wordpiece_m500_sru.py // test i-SRU + conv model
python test_wsj_wordpiece_m500_lstm.py // test LSTM model