This project addresses a machine learning task for voice-based access control, formulated as a binary classification problem in which audio recordings are transformed into spectrogram representations that serve as inputs to a neural network for speaker authorization. The work covers the full ML pipeline, including dataset compilation and preprocessing, data augmentation, spectrogram feature generation, and systematic experimentation with CNN design choices such as network depth, filter sizes, batch normalization, dropout, different optimizers, learning rate schedulers, and regularization, ultimately leading to the selection of a ResNet-18–based model as the best-performing solution.
Link to the dataset: https://www.kaggle.com/datasets/paulamatuszewska/spectrograms-aug/data