ESPnet extensions for semi-supervised end-to-end speech recognition

This repository contains evaluation scripts used in our paper

@inproceedings{Karita2018,
  author={Shigeki Karita and Shinji Watanabe and Tomoharu Iwata and Atsunori Ogawa and Marc Delcroix},
  title={Semi-Supervised End-to-End Speech Recognition},
  year=2018,
  booktitle={Proc. Interspeech 2018},
  pages={2--6},
  doi={10.21437/Interspeech.2018-1746},
  url={http://dx.doi.org/10.21437/Interspeech.2018-1746}
}

Full PDF is available in https://www.isca-speech.org/archive/Interspeech_2018/abstracts/1746.html.

how to setup

$ git clone https://github.com/nttcslab-sp/espnet-semisupervised --recursive
$ cd espnet-semisupervised/espnet/tools; make PYTHON_VERSION=3 -f conda.mk
$ cd ../..
$ ./run.sh --gpu 0 --wsj0 <your-wsj0-path> --wsj1 <your-wsj1-path>

NOTE: you need to install pytorch 0.3.1.

scripts

in root dir

• run.sh : end-to-end recipe for this experiment (do not forget to set –gpu 0 if you have that)
• sbatch.sh : slurm job script for sevaral pair/unpair data ratio and hyper parameter search (requires finished run_retrain_wsj.sh expdir for pretrained model params)

in shell/ dir

• show_results.sh : summarize CER/WER/SER from decoded results of dev93/test92 sets (usage: `show_results.sh exp/train_si84_xxx`)
• decode.sh : a script for decode and evaluate training model (usage: `decode.sh –expdir exp/train_si84_xxx`)
• debug.sh : we recommend to source debug.sh before using ipython to set path to everything you need

in python/ dir

• asr_train_loop_th.py : is a python script for initial-training with the paired dataset (train_si84)
• retrain_loop_th.py : is a python script for re-training with the unpaired dataset (train_si284)
• unsupervised_recog_th.py : is a python script for decoding by the re-trained model
• unsupervised.py : implements pytorch model for paired/unpaired learning
• results.py : implements chainer like reporter without chainer iterator used in training loop

results

train_set	dev93 Acc	dev93 CER	eval92 CER	dev93 WER	eval92 WER	dev93 SER	eval92 SER	path
train_si84 (7138, 15 hours)	77.6	25.4	15.8	61.9	44.2	99.8	98.5	exp/train_si84_blstmp_e6_subsample1_2_2_1_1_unit320_proj320_d1_unit300_location_aconvc10_aconvf100_mtlalpha0.5_adadelta_bs30_mli800_mlo150
+ train_si284 RNNLM		19.3	16.6	51.3	47.7	99.8	99.7	exp/rnnlm_train_si84_blstmp_e6_subsample1_2_2_1_1_unit320_proj320_d1_unit300_location_aconvc10_aconvf100_mtlalpha0.5_adadelta_bs30_mli800_mlo150_epochs15
+ unpaired train_si284 retrain	83.8	28.2	15.6	61.2	40.5	99.6	97.6	./exp/train_si84_retrain_None_alpha0.5_adadelta_lr1.0_bs30_el6_dl1_att_location_batch30_data_loss0.9
+ RNNLM		22.1	17.2	51.6	44.2	99.0	99.4	./exp/train_si84_retrain_None_alpha0.5_adadelta_lr1.0_bs30_el6_dl1_att_location_batch30_data_loss0.9/rnnlm0.1
+ unpaired train_si284 retrain w/ GAN-si84	83.5	26.3	15.0	59.9	40.0	99.4	97.3	exp/train_si84_paired_hidden_gan_alpha0.5_bnFalse_adadelta_lr1.0_bs30_el6_dl1_att_location_batch30_data_loss0.9_st0.5_train_si84_epochs15
+ unpaired train_si284 retrain w/ KL-si84	83.6	28.5	15.6	60.5	40.4	99.6	97.3	exp/train_si84_paired_hidden_gausslogdet_alpha0.5_bnFalse_adadelta_lr1.0_bs30_el6_dl1_att_location_batch30_data_loss0.9_st0.9_train_si84_epochs15
+ unpaired train_si284 retrain w/ GAN	84.2	22.1	17.9	50.9	44.2	99.2	99.4	./exp/train_si84_retrain84_gan_alpha0.5_adadelta_lr1.0_bs30_el6_dl1_att_location_batch30_data_loss0.9_st0.9_train_si84_iter5
+ RNNLM		22.1	17.9	50.9	44.2	99.2	99.4	./exp/train_si84_retrain84_gan_alpha0.5_adadelta_lr1.0_bs30_el6_dl1_att_location_batch30_data_loss0.9_st0.9_train_si84_iter5/rnnlm0.2
+ unpaired train_si284 retrain w/ KL	84.0	24.8	14.4	58.1	39.5	99.6	96.4	./exp/train_si84_ret3_gausslogdet_alpha0.5_bnFalse_adadelta_lr1.0_bs30_el6_dl1_att_location_batch30_data_loss0.9_st0.5_train_si84_epochs30
+ RNNLM		20.0	16.9	48.9	42.7	99.0	99.1	./exp/train_si84_retrain84_gausslogdet_alpha0.5_adadelta_lr1.0_bs30_el6_dl1_att_location_batch30_data_loss0.99_st0.99_train_si84/rnnlm0.2
+ unpaired train_si284 retrain w/ MMD	82.9	25.9	13.9	59.7	38.4	99.2	96.7	./exp/train_si84_ret3_mmd_alpha0.5_bnFalse_adadelta_lr1.0_bs30_el6_dl1_att_location_batch30_data_loss0.5_st0.99_train_si84_epochs30
train_si284 (37416 utt, 81 hours)	93.9	8.1	6.3	23.8	18.9	92.4	87.4	exp/train_si284_blstmp_e6_subsample1_2_2_1_1_unit320_proj320_d1_unit300_location_aconvc10_aconvf100_mtlalpha0.5_adadelta_bs30_mli800_mlo150
+ train_si284 RNNLM		7.9	6.1	22.7	18.3	89.7	84.1	./exp/rnnlm_train_si284_blstmp_e6_subsample1_2_2_1_1_unit320_proj320_d1_unit300_location_aconvc10_aconvf100_mtlalpha0.5_adadelta_bs30_mli800_mlo150_epochs15

• Acc: character accuracy during training with forced decoding
• CER: character error rate (edit distance based error)
• WER: word error rate (edit distance based error)
• SER: sentence error rate (exact match error)
• all the exp path starts with exp/... is placed to /nfs/kswork/kishin/karita/experiments/espnet-unspervised/egs/wsj/unsupervised on NTT ks-servers

smaller paired train data results

contact

email: [email protected]