Description
Many audio anomaly-detection systems (e.g., for machine condition monitoring, automotive acoustics, or environmental sound monitoring) rely on time-dependent representations and must remain robust under changing operating conditions. Liquid Neural Networks—especially Closed-form Continuous-time models (CfC)—model temporal dynamics in continuous time and can be compute-efficient, adaptive, and well-suited for Edge AI. This thesis investigates the performance of CfC-based models in anomalous sound detection, for example via reconstruction error (autoencoder) or prediction error (time-series forecasting), and compares them to common sequence baselines (e.g., LSTM/GRU/TCN/Transformer), including an analysis of robustness, generalization, and inference cost.

Task
Develop and evaluate a CfC-based method for audio anomaly detection, including an appropriate feature pipeline (e.g., log-mel/MFCC/embeddings), a downstream formulation (forecasting or autoencoder reconstruction), and a thorough comparison to strong baselines. Analyze performance, robustness to domain shifts/noise, and computational efficiency with an Edge-AI perspective.

Prerequisites

• Strong Python and PyTorch skills
• Fundamentals of audio/signal processing (STFT, mel-spectrograms, feature design)
• Interest in deep learning
• Basic understanding of sequence modeling (RNN/TCN/Transformer) and experimental design