Hierarchical Attention Mechanism for Domain-invariant Spoken Language Identification via Inter-domain Alignment and Intra-domain Language Discrimination

Modern methods for spoken language identification (LID) have demonstrated promising results when trained on large datasets. However, their effectiveness is considerably impacted by the discrepancies between the training and testing distributions. Unsupervised domain adaptation (UDA) addresses the domain discrepancies by aligning feature distributions across source and target domains without using labeled target data, thereby ensuring the model performs consistently across both domains. This alignment, however, may come at the cost of reduced model discriminability. To enhance performance under domainmismatched conditions, a LID system should learn domaininvariant representations of speech while maintaining language discriminability at every level, from initial segment-level representations to final utterance-level representations. To address this issue, in this paper, we propose a hierarchical attention network that simultaneously enhances model domain invariance through inter-domain alignment and improves model's language discriminability using intra-domain contrastive discrimination. The performance on the target domain shows that the proposed method exceeds the effectiveness of the top-performing existing approaches.