Cross-corpora spoken language identification with domain diversification and generalization

This work addresses the cross-corpora generalization issue for low-resourced spoken language identification (LID) problem. We have conducted experiments in context of Indian LID and identified strikingly poor due to corpora-dependent non-lingual biases. Our contribution this is twofold. First, we propose domain diversification, which diversifies limited training data using different audio augmentation methods. then concept maximally diversity-aware cascaded augmentations optimize fold-factor effective diversification data. Second, introduce idea considering methods as pseudo-domains. Towards this, investigate both domain-invariant domain-aware approaches. system based on state-of-the-art emphasized channel attention, propagation, aggregation time delay neural network (ECAPA-TDNN) architecture. extensive with three widely used corpora research. In addition, conduct a final blind evaluation our proposed subset VoxLingua107 corpus collected wild. demonstrate that more promising over commonly simple The study also reveals solution than diversification. notice learning performs better same-corpora LID, whereas suitable generalization. Compared basic ECAPA-TDNN, its extensions improve EER up 5.23%. contrast, performance test scenarios.