Unsupervised Feature-Space Domain Adaptation applied for Audio Classification

Domain adaptation is a fundamental technique to ensure that deep neural networks perform robustly even in unknown target domains. In this paper, we study Z-Score normalization, relaxed instance frequency-wise normalization (RFN), and feature projection-based DA (FPDA) for unsupervised feature-based domain adaptation. With a focus on acoustic monitoring, we investigate the classification of individual sounds and acoustic scenes as the main use cases. Based on a systematic study of different normalization techniques and data partitioning strategies, our results confirm that an individual normalization per frequency band is beneficial for sound classification, whereas a global classification applied to individual data instances is beneficial for acoustic scene classification. As another main contribution, we propose the IFPDA method, essentially, is a variation of the original FPDA configuration, allowing it to be applied independently to each instance, and results in a substantial performance improvement and even outperforms all other normalization methods in the acoustic scene classification task.