Enhancing Indian Language Speech Recognition Systems with Language-Independent Phonetic Script: An Experimental Exploration

Abstract

India has a distinct linguistic profile that makes it extremely challenging to train Automatic Speech Recognition (ASR) systems correctly because most Indian languages have limited training material and are generally low-resource in nature. However, due to their similarity in phonemes, these languages offer an opportunity to create a single speech recognition technology. This study proposes an optimized phonetic script which can be generalized for all major Indian languages. To prove the efficiency of the phonetic script, speech recognition models using phonetic and language scripts in Hindi and Malayalam language were created using a wave2Vec2-based Deep Neural Network (DNN) model via transfer learning. Furthermore, a model based on Long Short-Term Memory (LSTM) is created to translate phonetic script text back into its original languages. The findings show that the phonetic script ASR model performed noticeably better than the language-specific model, reducing WER roughly to 2%, especially for the Hindi language, which is further reduced up to 1% for the model trained with mixed language. This demonstrates the model's ability to improve performance by using cross-lingual phonetic similarities. This study establishes the foundation for cross-linguistic, scalable ASR systems that use phonetic similarities to enhance ASR performance in low-resource language contexts across India.