Bridging Domain Gaps With ProtoAlign: Teacher–Student Few-Shot Prototype Alignment for Cross-Domain Spoken Language Recognition

Spoken language recognition in low-resource settings is hindered by domain shift and limited labeled data.We propose ProtoAlign, a teacher–student few-shot prototype alignment framework that learns domain-invariant, language-discriminative representations with minimal target supervision. The student uses a compact transformer-style backbone with Feature Reweighting Layer (FRL). Source-domain class prototypes are maintained as exponential moving averages and serve as stable anchors. A target-to-source Information Noise Contrastive Estimate(InfoNCE) alignment term pulls few-shot target embeddings toward their language-matched source prototypes, while a lightweight knowledge-distillation loss from a source-only teacher preserves source accuracy. Warm-start schedules for the alignment and distillation weights stabilize optimization, and a pairing sampler ensures each batch contains target samples with same-language source counterparts. We evaluate on five Indian languages across five heterogeneous domains (All India Radio (AIR), Common Voice (CV), Kaggle, Indian Institute of Technology Hyderabad (IIT-H)) and Indic TTS. With at most ten labeled target examples per language, ProtoAlign consistently outperforms a strong transformer baseline in cross-domain tests and produces visibly tighter, more domain-invariant clusters in the embedding space. These results indicate that prototype anchoring combined with gentle teacher guidance provides a simple, scalable, and label-efficient path to robust cross-domain spoken language recognition.