/speech_recognition_translation

Primary LanguagePython

speech_recognition_translation

1. Introduction

There are two function module in this project. One for speech recognition, which adopts Google WaveNet DNN, the other for machine translation, which adopts seq2seq + attention DNN.

2. Environment

  • python2.7
  • tensorflow 1.0.0
  • tflearn 0.3.2
  • numpy 1.13.3
  • six 1.11.0

3. Usage

Download the VCTK corpus and uncompress it to ./data/asr/. This corpus is used for training speech recognition model. Since it is an English corpus, the model will be able to recognise English speech. To train the speech recognition model:

python train_asr.py

To recognise a speech:

python recognise.py

The data for training machine translator has been put in ./data/nmt/. They are some subtitles of TED speech. Since it is an English-French corpus, the model can only translate English to French. To train the neural translation model:

python train_nmt.py

To translate a text:

python translate.py

4. TODO

  • do fine tuning for the model
  • add BLEU score to check the model performance
  • add beam search to NMT

5. Reference