An LSTM Approach to Language Identification Across Indian Languages Using MFCC

Spoken language identification aims to identify the language of an audio. As India is a multilingual country, it is important to recognize the language. This work proposes an approach based on Mel Frequency Cepstral Coefficients (MFCC) for feature extraction and Long Short-Term Memory (LSTM) for identification of ten languages: Hindi, Kannada, Bengali, Telugu, Tamil, Marathi, Urdu, Malayalam, Punjabi, and Gujarati. There are a few audio databases available for Indian languages. The experiment uses the "Audio Dataset with 10 Indian languages," where the audio length is around 5 secs. This work evaluates the accuracy of deep learning models in language identification and compares the results. The proposed model achieves an accuracy of 88%, considered a strong performance for language identification.