Voice pathology detection

This is my bachelor thesis project. Voice pathology detection is a patient condition diagnostics system based on Machine Learning algorithms. Vanilla version was based on logistic regression and random forest, though when I discovered CNN's and computer vision I thought it would be nice to test a deep learning approach. To conclude the project contains approaches to voice pathology detection problem:

• Classic machine learning approach, two algorithms logistic regression and random forest with overall accuracy ~70%,
• Convolutional neural network utilized to classified patients based on extracted spectrograms, overall accuracy ~70%.

Responsibilities:

• Built machine learning project pipeline: data analysis, features creation, model preparation, validation

Installation

Just clone the repository. Required dependencies are given in requirements.txt file:

pip install -r requirements.txt

Classifiaction

The classification workflow is presented in Jupyter Notebooks:

• classic machine learning voice_pathology/classic_ml/classification.ipynb
• CNN voice_pathology/cnn/classification.ipynb

Dataset

Classification is based on Saarbruecken Voice Database a collection of voice recordings from more than 2000 persons. There are two Jupyter Notebooks presenting data preparation:

• classic machine learning voice_pathology/classic_ml/audio_analysis.ipynb
• CNN voice_pathology/cnn/audio_analysis.ipynb