AASP-processing

This reposistory contains scripts to train speech prosody classifiers, to be used in the Automatic Analysis of Speech Prosody application. The scripts collected here can be used to use audio features extracted with the speech prosody analysis tool AuToBI to predict human annotations in .TextGrid format.

For more documentation about the file types: - arff format, as used by the machine learning library Weka. These files are a specialized version of a csv, in which each feature name is listed as @ATTRIBUTE on top of the file, and the feature value is a cell in a comma-separated grid - TextGrid format, as used by audio annotation and analysis software Praat.

Prerequesites

1. Make sure you have Python 3.6 installed (e.g. via Anaconda)
2. If you do not want to install the required libraries globally, make a conda environment (see instructions here)
3. Navigate to this directory and run pip install -r requirements.txt

Running

The repository contains two scripts: 1. process_arff can be used to combine audio features computed by AuToBI (in arff format) with human annotations (in TextGrid format). python process_arff.py -a PATH/TO/YOUR/ARFF/FEATURE/FILES -t PATH/TO/YOUR/TEXTGRID/FILES will: - adjust the headers (@ATTRIBUTEs) of the arff feature files and write new files into a directory. By default, this will be subdirectory called 'processed' in the source file directory, you can set another directory with the -c flag. - step through the directory with TextGrid files (including subdirectories) and collect the annotated intonations, and combine them with the features from the arff files in a large output file. By default, the complete table is written to combined.arff, and a table with only numerical parameters will be written to combined_nostring.arff in this directory. You can set another filename with the -o flag. - For help with the script, call python process_arff.py --help 2. As a next step, train_classifiers documents how to train a new classifier using Python's scikit-learn, in which the features extracted from the .arff file will be use to predict the annotations extracted from the .TextGrid file. It shows how to achieve train / test split, and suggests that a random forest classifier may achieve around 0.9 accuracy for prediction of tone and accent, which were the annotated features in data used for making a proof of concept.