Prodylab-Alignment adapters and other parsers

Adapters to adapt the danpass corpus data to fit with the prosodylab-alignment software.

Everything is implemented using python version 3.4.3.

If the training is being made on a ubuntu linux server, then the server can be setup with the script setupServer.py. Just transfer the project to the server, then run: python3 serverSetup.y This should install the necessary dependencies, including downloading prosodylab from github. This has been tested launching Amazon AWS instances. HTK-3.4.1 should be available in .tar.gz or .tar form in the parent directory of the directory this is being run in. This script will the unzip HTK and install it.

In order to set up the project call: python3 setupTraining.py -a "path to aligner" -d "path to danpass corpus" -p "training set size in percent" Example: python3 setupTraining.py -a Prosodylab-Aligner/ -d DanPass/ -p 60

After setting up the project, the prosodylab aligner can be used to train on the files in directories/MonoTrain/, directories/DialTrain or directories/AllTrain. Dictionary and .yaml file will be in directories/Parameters.

To evaluate performance run: python3 setupEvaluation.py -f "path to danpass corpus"

To fix .TextGrid files in a selected tier. Fixes danish characters and fixes some white space differences. (Creates new fixed files, does not overwrite old) python3 setupEvaluation.py -d "Path to directory with TextGrid files" "tier name"

For two directories compare performance between all .TextGrid files with the same names. Always compares the first tier. If run on cleaned .TextGrid directory then there will only be one tier. python3 setupEvaluation.py -c "path to directory one" "path to directory two"

The file createDialogueGrid.py will create a file for each dialogues file where every utterance (uninterrupted sequence of word by a speaker) is transcribed with timings and who uttered it. To run it: python3 createDialogueGrid.py -d "Path to folder with dialogue files" -c "path to danpass corpus"

Prosodylab uses HTK which has an issue with large files. The script splitWav.py can be used to mitigate this problem: python3 splitWav.py -d "path to danpass corpus" -mono "path to monologue sound files" -dial "path to dialogue sound files" -p "size in percent of training set"

This will split the sound files into files of size approximately 5Mb.

The script crossValidation.py can setup for k-fold cross validation: python3 crossValidation.py -k "k in k-fold" -m "path to folder with monologues .lab and .wav files" -d "path to folder with dialogue .lab and .wav files" This then creates two subdirectories inside "directories" called crossValidationTest and crossValidationTraining where each of these subdirectories will have subdirectories called "roundi", where 1 <= i <= k.