Large Language Model for Kannada Speech Recognition
G Hemakumar, Dr. Pushpalatha M, Nagandra Nath Giri, Santhoshkumar B.N · International Journal of Engineering Development and Research · 2025
this paper discuss on designing the large language model for Kannada automatic speech recognition system. We are replacing ASR language model with the KLLM (Kannada Large Language Model) for the enhancement of recognition rate of Kannada dialect speech, whispered speech, noise speech signals etc.,. The Kannada language has a SOV (Subject-Object-Verb) structure and rich system of inflections and agglutination. The Large Language Models (LLMs) are advanced artificial intelligence (AI) systems designed to process and generate human-like text. LLM is capable of understanding and generating human-like text. Hence, we have built using deep learning techniques (neural network), particularly transformer-based architectures, and are trained on 150 million Kannada textual data. Algorithm designed for recognition of speaker independent continuous Kannada speech made by different dialects speakers. The novelty of algorithm is that implementing LLM for ASR system for self-supervised learning to handle multiple Kannada dialects speaker’s speech recorded by mini-microphone, headphone and cell phones in natural environment. This create the robustness of the algorithm in handling noisy waves, whispered waves and also work for out-of-vocabulary (OOV) words. In this proposed system the WRR is about 91.04% and WER is 8.96%. All computations made using python programming language.