Tamil Natural Language Voice Classification using Recurrent Neural Networks

Abstract

Audio Classification systems are used to classify the given audio into N outputs. Various models can accurately classify a given sound. But few can accurately classify a given Natural Language (Tamil) Voice, especially Tamil Vowels pronounced by Children with learning disabilities. Voice classification is done by recording the audio and converting it into digital data. The audio samples then undergo a feature extraction process to extra Mel Frequency Cepstral Coefficients. These coefficients are then used as input to the Deep Recurrent Neural Networks (DRNN) (LSTM) to accurately classify them into Tamil vowels. This paper focuses on a learning system (android app), with an on-device inference model developed using TensorFlow Lite, that records the children’s voice data, classifies them, and provides accuracy and training to improve their pronunciation.