Data augmentation for speech separation

Deep learning models have advanced the state of art monaural speech separation. However, performance a separation model considerably decreases when tested on unseen speakers and noisy conditions. Separation trained with data augmentation generalize better to In this paper, we conduct comprehensive survey techniques apply them improve generalization time-domain models. The include seven source-preserving approaches (Gaussian noise, Gain, Time masking, frequency Short stretch, Pitch shift) three non-source preserving (Dynamix mixing, Mixup, Cutmix). After hyperparameter search for each method, test augmented by cross-corpus testing datasets (LibriMix, TIMIT, VCTK), identify best combination that enhances generalization. Experimental results indicate several strategies (CutMix, Dynamic mixing) resulted in performance. Finally, combinations also improved even fewer training are available.