Course project on continuous audio signal monitoring including collection, processing, and authentication. Deep Learning approach for audio-based user authentication.
The goal of this project is to utilize mel coefficient features and deep learning for user authentication by voice analysis. We implement various temporal-spectral features with no prior knowledge and learn the best features for this task:
- Collect sound data with the data_acq_gui script.
- Perform feature analysis to analyze quality of various sound features
- Build a CNN model optimizing all possible combinations of features to detect the best feature combination
- Train a CNN and save weights
- Evaluate the CNN with ROC and PRC curves, measure AUC, and evaluate various working points (max sensitivity, max specificity, max accuracy)