PAG_recipe.md

Recipe to create Pre-alignment guides for training

Step 1 - Install tools

cd tools

bash compile_festival.sh

If you dont already have an account with HTS, you need to create one in order to download HTK. Get an account here: htk as you will need a username and password for the next step

bash compile_htk.sh $username$password

To test HTK and Festival is installed and working, the following should run without error:

HCopy

festival

Step 2 - Get data

mkdir data

cd data

wget link to data

cp -f train.csv {data_path}

Step 3 - Edit hparams.py

wav_path = 'data/wavs_train/'
data_path = 'data/CB_JE'
book_names = ['CB-JE']
voc_model_id = 'blizzard_vocoder'
tts_model_id = 'blizzard_baseline_JE'
metadata = "train.csv"

Step 4 - Extract features

python preprocess --hp_file hp_JE.py

Step 5a - Run forced alignment

python forced_alignment.py --hp_file hp_JE.py

This will output all the necessary files for forced alignment in {data_path}/labels

Step 5b - Run alignment on different model (optional)

Since we only using 1 book in this recipe, it will be better to train the aligner with all the data. In this case, you need to run Step 5a with all the data and then run:

python scripts/align_data.py --lab_dir {labels folder of JE book} --model_dir {labels folder of all books}

The output label files will be in {data_path}/labels/label_state_align

Step 6 - Get durations and create guides

python get_durations.py --hp_file hp_JE.py

The output guides will be saved in {hp.data_path}/attention_guides

Step 7 - Train PAG

python train_pag.py --hp_file hp_JE.py