ANALYZING LANGUAGE IN MULTILINGUAL SPEECH USING DEEP NEURAL NETWORK

Language recognition is the process of determining the language spoken. Motivated by the impressive gain in the performance of language recognition, we adapt the deep neural networks to the problem of language recognition analysis. In the prior work we consider the application of education institute, where teachers comprehend their speech in multiple languages. We then analyze the different aspects of primary (English) and secondary (Hindi, Gujarati) languages spoken. We have prepared a basic work flow of the proposed solution. Speech features are modelled by MFCC parameterization method. In the current work we have used Long Short-Term Memory (LSTM) neural network as our deep neural network model. Results are carried out on Indic TTS dataset provided by IIT Madras. Our result shows that Deep Neural Network (DNN) gives evident results especially when the amount of training data is more. Analysis is done considering various cases 1) 1 hidden layer - 4 hidden layer LSTM network 2) 10s – 15s audio 3) small – large dataset. The analysis carried out can be helpful at administration level to measure the quality of teaching in a class which can lead to improvement in the education system.