Attention-based sequence to sequence AST System. The model uses ASR and MT Systems for initialization (Transfer Learning).

The project is based on the code of Alexandre Bérard https://github.com/eske/seq2seq

## Getting Started

### Prerequisites and Installing

1. Follow the instructions https://github.com/eske/seq2seq to install the dependencies and the seq2seq system.
2. Download the folder cluster/home/proj/speechrecognition/TRANSFER_LEARNING_AST 

### Input Data

* Used Input Data (given):

* Used pre-trained ASR and MT Models (given)

ASR ("best-276000" was used for initialisation): 

MT ("best-98000" was used for initialisation): 

## Learn the system
## READY-TO-USE MODEL:

 **Our ready-to-use best model:** 

* Other checkpoints and log files: 

--- First 220000 steps (srarting from 98000 step):

--- Next steps:

* Configuration files:

--- First 220000 steps (starting from step 98000):

--- Next steps:

## TRAIN YOUR OWN MODEL:

- Create new output folder, change directions/hyperparameters/checkpoints in transfer_1.yaml configuration file.

- Change to your seq2seq folder, run the following script:

...where CONFIG is the direction to the transfer_1.yaml file on your computer.


## Other possobilities:

* Input Data produced by our group:
tar audio files:

npz, vocab und spain textfiles:

* To generate your own input Data with scripts of our group:
cluster/home/proj/speechrecognition/seq2seq

* Use other versions of pre-trained ASR and MT Models (new ones):

## Built With

* [Alexandre Bérard] https://github.com/eske/seq2seq) 

## Authors

* **Chernenko Tetyana (Tatjana)** 
* **Liang Siting**
{chernenko/liang}@cl.uni-heidelberg.de