/zr2019

Primary LanguagePython

########## BASELINE FOR ZEROSPEECH 2019#####

This baseline works as two independant systems.

First it trains BEER (https://github.com/beer-asr/beer) an acoustic unit discovery algorithm. From a dataset of audio files, it learns an unsupervised set of acoustic units.
These phones are then fed into OSSIAN (https://github.com/CSTR-Edinburgh/Ossian), a text-to-speech pipeline. From sentences, written with units from BEER, OSSIAN learns to synthesize speech. 

##########HOW TO TRAIN THE BASELINE AND BUILD A SUBMISSION##########

# All dependancies and systems are pre installed in the docker. You can check what was installed by reading set up scripts at: 

vim $HOME/baseline/set_up_baseline.sh
vim $HOME/eval/set_up_eval.sh

# Train on toy dataset:
cd $HOME/baseline/training
bash train_baseline.sh /home/zs2019/baseline /home/zs2019/databases/english_small/ 1
# Train on full dataset
bash train_baseline.sh /home/zs2019/baseline /home/zs2019/databases/english/ 1

# Decode sentences and make a submission from the baseline
bash decode_and_make_submission.sh /home/zs2019 submission


##########HOW TO EVALUATE A SUBMISSION##############################

# A submission for the english database should be structured like this :

$HOME/submission:
	metadata
	code/*
	english/test/*.txt and *.wav

# You need to validate your submission before evaluating it
cd $HOME/tools/
bash validate.sh $HOME/baseline/training/submission/zip english

# You can then run the evaluation script (note that the evaluation script will also do the validation by default)
cd $HOME/tools/eval
bash evaluate_submission.sh  /home/zs2019/ /home/zs2019/submission