recipe.sh demonstrates an end-to-end usage of this codebase for experiments on cca-mel, mi-phone, and cca-phone.
The for ... loops in the script are written so only to give the user an idea of what constitutes a complete workflow. The script will run fine as is but will be very slow to execute as everything is processed serially. The individual commands within the loops can be and are encouraged to be run parallely. The steps themselves should be executed in the ordered fashion, i.e. step 2 before step 3, step 3 before step 4.
Before running the recipe:
- Make sure that the paths are set correctly to reflect
- Librispeech data location (
path_to_librispeech_data) - Location where you want the alignment files to be saved(
alignment_data_dir) -- it takes 6Gb of space - Location where the extracted representation files should be saved (
save_dir_pth) -- the representations extracted form each dev-clean sample set take upto 12Gb and for each train-clean sample set 23Gb.
- You can change the
stepsvariable to skip the steps previously done. For example, if you already have the alignment files saved atalignment_data_dir, setsteps=2so that the script skips that step and goes over to the data sampling and extraction step (step 2)