Harnessing XLSR for Spoken Language Identification in Indian Languages

Abstract

Spoken Language Identification (SLID) is an important part of multilingual speech systems, especially in India, where Indo-Aryan and Dravidian languages exist alongside each other and have many phonetic similarities. These similarities, along with accent variation and limited labeled data, create challenges for traditional SLID systems. Handcrafted acoustic features like Mel Frequency Cepstral Coefficients (MFCC) capture short-term spectral information but often struggle to tell apart closely related languages in real-world situations. This paper aims to assess how pretrained transfer learning models impact spoken language identification by using cross lingual knowledge gained from large multilingual speech data. To do this, we conduct a comparative analysis between an MFCC model and one based on Cross-Lingual Speech Representations (XLSR). Both models are evaluated using a deep neural network classifier on a multilingual Indian speech dataset. The results show that the MFCC-based approach has lower accuracy, while the pretrained XLSR-based model greatly improves language identification performance by capturing contextual, phonetic, and language-invariant speech traits from raw audio. The findings indicate that pretrained transfer learning offers a strong and scalable solution for spoken language identification, especially in multilingual settings with languages that are closely related. This underscores the value of using cross-lingual pretrained speech models to create robust and effective SLID systems for the future.