modeling

models

Estimate the prosodic dissimilarity of utterances

To start the application, run the Python script gui_main.py.

The GUI application consists of three main views: the "To fragment" view, the "To mean" view, and the "Overall" view.

The "To fragment" view lets you search for and select a fragment by ID, and then displays the fragments that are most similar and dissimilar according to the similarity measure. Next to a fragment is a button to play its audio and a field to enter notes on observations. This view displays fragments only in the same language, as the models are designed to compare fragments within the same language.

Create feature correlation figures

To create feature correlation figures, run make_feature_corrs_fig.py. Below is the feature-by-feature correlation figure for English versus Spanish DRAL 4.0 short fragments.

Model prosody mappings

1. linear regression model (model_linear_regression.py)
2. k-nearest neighbors regression (model_knn.py)
3. shallow feed forward neural network (neural_network.py)
4. naive baseline (model_naive_baseline.py)
5. synthesis baseline (model_synthesis_baseline.py)