Abstract: We propose a novel approach for language identification, specifically tailored for the challenging task of distinguishing homophonic short utterances. Homophonic utterances, where different languages produce similar sounds, pose a significant challenge in multilingual speech processing. We introduce a Convolutional Neural Network (CNN) architecture optimized for extracting discriminative features from audio segments. These homophonic utterances, characterized by similar sounds across different languages, are notoriously difficult to distinguish and thus require specialized techniques in multilingual speech processing.
Experimental results demonstrate the superiority of our CNN-based approach in language identification, making it a valuable contribution to the field of multilingual speech processing. The experimental study was carried out on a real-time dataset comprising Hindi, Kannada, Telugu, Marathi, and several other languages. In addition to the CNN-based approach, we also employed three traditional classifiers: Deep Learning, Convolutional Neural Network, and others. Experimental evaluations underscore the effectiveness of the CNN-based approach, showcasing its ability to achieve impressive accuracy in identifying languages within homophonic contexts. To provide a comprehensive assessment, we implemented approaches for different duration intervals, including 5 seconds, 10 seconds, and 20 seconds.
This innovative methodology addresses a pressing challenge in language identification, particularly in the context of homophonic utterances, and offers a promising solution for multilingual speech processing. Through rigorous experimentation and comparative analysis, our approach demonstrates notable advancements in accuracy and performance, thereby contributing significantly to the field of language identification and multilingual speech processing.
Keywords: Multilingual speech processing, CNN, Real-time datasets Hindi, Kannada, Telugu, Marathi, tamil ,Urdu, 5 sec, 10 sec, 20 sec, Deep learning.
|
DOI:
10.17148/IJARCCE.2024.134212
