Leveraging Augmented Samples as an Auxiliary Task for Enhancing Spoken Language Identification in Indian Languages

Spoken Language Identification (SPLID) in the context of Indian languages presents unique challenges due to the significant overlap of linguistic features within the same language family. Discriminating between closely related languages is particularly difficult. While deep learning-based SPLID systems have demonstrated substantial performance gains through augmented samples, these augmented synthetic data can sometimes lack natural voice modulation, exacerbating the challenge of identifying languages from the same family. To address this, we propose an orthogonal auxiliary task—actual vs. spoof identification—within a multitask learning (MTL) framework. By assigning higher weightage to real data and employing augmented samples cautiously, we aim to enhance the model’s ability to distinguish between closely related languages. Our findings suggest that including this auxiliary task in the MTL paradigm improves classification accuracy compared to single-task models.