A Deep Learning-Based Study on Inter-Language Confusion Assessment Using Spoken Language Identification for Indian Languages

This study investigates the effectiveness of various features and state-of-the-art spoken language identification (LID) systems in quantitatively analyzing the classification performances and confusions across languages. We focus on 23 Indian languages from the IIITH-ILSC corpus for our analysis. Initially, we employ conventional mel-frequency cepstral coefficients (MFCCs) with the ECAPA-TDNN classifier. To enhance classification performance, we further implement self-supervised learning (SSL)-based features derived from the transformer layers of the wav2vec 2.0 XLSR model, observing a substantial improvement in accuracy, (90.73%) compared to the MFCC-based model (72.86%). We further expand our approach by applying wav2vec 2.0 features to a recently proposed conformer-based architecture. For all these state-of-the-art systems, we then compare the language confusion patterns to check if the most confusing language pairs remain consistent. To further examine whether these obtained confusion patterns are framework-dependent, we perform comparative analyses among LID frameworks with multiple best-performing checkpoints and random initializations. To our knowledge, using the pre-trained self-supervised learning (SSL)-based representations and large-scale DNN-based classifiers, this is one of the first studies that explicitly explore the interpretability of LID frameworks.