Multilingual and Mixed-lingual Digit Speech Recognition System for Indian Context
S. Lalitha, N Rachana, Vinay Bhargav J · 2023
Voice assistants, customer service center automation, and voice-controlled devices all rely heavily on digit speech recognition. For multilingual countries like India, Canada, and Belgium to communicate in more than one language, there is a need for multilingual and mixed-lingual digit speech recognition (DSR) systems. This is the main emphasis of the proposed work. An extensive database of digit utterances from 0 to 9 for various Indian languages of Telugu, Kannada, Hindi & Tamil are self-recorded and used for the experimental work. Following the initial preprocessing by noise elimination from the self-recorded samples, spectrographic characteristics were extracted, and finally, a Convolutional Neural Network (CNN) is applied for classification to identify the digits. The proposed multilingual DSR model resulted in an accuracy of range 89% to 96% and a loss value between 0.144 to 0.878 varying for each language whereas the mixed-lingual DSR model resulted in an accuracy of 92% and a loss value of 0.615.