Automatic Prediction of speech intelligibility using speaker embeddings

Two types of embeddings may be used in the following system:
	X-Vectors   (via Hugging Face) dim:512
	ECAPA_TDNN  (via Hugging Face) dim:192

The system designed makes use of a Shallow Neural Network whose first 
dimension is conditioned on the size of the embedding used.

The model architecture can be found on ./models/model.py

The full list of parameters and path to files used can be found on the
config file (see ./configs/parameters.yaml)

the script ./dataloader/embedding_extract.py provides a function to extract
the speaker embeddings. See config file for PATH setup for .wav and .pickle
files.

Input TRAIN, TEST and VALIDATION .csv files should be in the following form:

	embedding_file.pickle,SEV,INT,V,R,PD

	SEV - Speech Disorder Severity
	INT - Speech Intelligibility
	V - Voice Quality Analysis
	R - Resonance
	P - Prosody
	PD - Phonemic Distortions

train.py script trains and validates the system (see train and validation file
path in the configs file)

test.py script tests the system (see test file in the configs file)

To directly predict speech intelligibility using a pretrained model go to ./predictor
and use the method "predict" of the class PythonPredictor (in file predictor.py)

Example:
	$python3
	>> import predictor
	>> pred = predictor.PythonPredictor()
	>> pred.predict('PATH_TO_WAV_FILE')