Sequence to sequence with noise injection

Overview

Experiments

Experiment 1

Experimental procedure

• 1. Prepare GPSRDATA
• 2. Make "original" Dataset

  ./create_original_dataset.sh

  • original_dataset
    • test.csv dev.csv train.csv
• 3. Make "edited" Dataset (Ineject Noise)

  learning_dataset

  • train/rho=<noise_level>_edited_phoneme.in
  • train/original_sentence.in
  • train/phoneme.in
  • train/label.out
  • dev/rho=<noise_level>_edited_phoneme.in
  • dev/original_sentence.in
  • dev/phoneme.in
  • dev/label.out
• 4. Prepare Utterance Data
  • 1. Speech Recognition

    ./SpeechRecognition.sh

• 5. Learning

python -m nmt.nmt --src=in --tgt=out\
--vocab_prefix=/path/to/Dataset/Vocab\
--train_prefix=/path/to/Dataset/traindata\
--dev_prefix=/path/to/Dataset/devdata\
--test_prefix=/path/to/Dataset/testdata\
--out_dir=/path/to/Noisy_model\
--num_train_steps=12000 --steps_per_stats=100 --num_layers=1 --num_units=128 --dropout=0.2 --metrics=bleu\
--attention_architecture=standard --attention=luong`

• 6. Inference

condition

• 隠れ層 : 1層,128次元
• ドロップアウト確率 : 0.2

Experiment 2

Experimental procedure

condition

Reference

• Yuuki Tada, Yoshinobu Hagiwara, Hiroki Tanaka and Tadahiro Taniguchi. Robust Understanding of Robot-Directed Speech Commands Using Sequence to Sequence With Noise Injection, 2020.

LINK

• https://github.com/komeisugiura/GPSRsentence_generator
• https://github.com/tensorflow/nmt

Main Contributors

• Yuki Tada (ex-student of Ritsumeikan University)

maintainor

• Hiroki Tanaka (assistant researcher of Ritsumeikan University)