Scholarlink

<p>Spoken language identification (SLID) under real-world class imbalance remains a bottleneck for multilingual voice applications. This study introduces an incremental, imbalance-aware framework that fuses self-supervised XLS-R embeddings through a dual-stream gated-attention classifier and augments minority classes with diffusion-based audio synthesis. Class-balanced focal loss and elastic-weight consolidation jointly preserve recall for low-resource languages while preventing catastrophic forgetting when new languages are added. On the Indian languages audio dataset, the framework attains a macro-<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\hbox {F}_1\)</EquationSource> </InlineEquation> of 0.995 for three languages and sustains 0.818 after expansion to ten languages, surpassing the strongest published baseline by 6 percentage points. A Friedman aligned-rank test yields <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(\chi ^2(3)=0.067\)</EquationSource> </InlineEquation>, <InlineEquation ID="IEq3"> <EquationSource Format="TEX">\(p=0.9955\)</EquationSource> </InlineEquation>, and a critical difference of 2.708 at <InlineEquation ID="IEq4"> <EquationSource Format="TEX">\(\alpha =0.05\)</EquationSource> </InlineEquation>, indicating no significant differences among the compared configurations. These results demonstrate the framework’s robustness under severe class imbalance and its suitability for scalable, low-resource SLID deployment in linguistically diverse environments.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Incremental imbalance-aware deep learning framework for multilingual spoken language identification

  • 期刊论文

  • 发表日期：

    2026年1月6日

查看全文
  • Vishakha Tomar,
  • Shubhra Dixit,
  • Pardeep Sangwan

摘要

Spoken language identification (SLID) under real-world class imbalance remains a bottleneck for multilingual voice applications. This study introduces an incremental, imbalance-aware framework that fuses self-supervised XLS-R embeddings through a dual-stream gated-attention classifier and augments minority classes with diffusion-based audio synthesis. Class-balanced focal loss and elastic-weight consolidation jointly preserve recall for low-resource languages while preventing catastrophic forgetting when new languages are added. On the Indian languages audio dataset, the framework attains a macro- \(\hbox {F}_1\) of 0.995 for three languages and sustains 0.818 after expansion to ten languages, surpassing the strongest published baseline by 6 percentage points. A Friedman aligned-rank test yields \(\chi ^2(3)=0.067\) , \(p=0.9955\) , and a critical difference of 2.708 at \(\alpha =0.05\) , indicating no significant differences among the compared configurations. These results demonstrate the framework’s robustness under severe class imbalance and its suitability for scalable, low-resource SLID deployment in linguistically diverse environments.