INDIAN SPOKEN LANGUAGE IDENTIFICATION USING CONVOLUTIONAL NEURAL NETWORK (CNN)

Abstract

A language identification system's goal is to determine the language used in a given spoken utterance. The methodology for language identification might be either implicit or explicit, depending on the availability of comparable text for a specific utterance. We only have accessibility to the acoustic signal in an implicit situation, and statistical models are formed over the features derived from the acoustic signal. In an explicit scenario, the speech has a corresponding text, and norm algorithms are employed to recognize the language. For language identification, this thesis provides an implicit technique based on Convolutional neural network (CNN) models. For the task of language identification, CNN models have recently been presented. CNN models were investigated in this thesis in the Indian context. In this work, the IIIT-H Indic speech database is used for carrying out the experiments. This database contains speech samples of 7 Indian spoken languages such as Bengali, Hindi, Kannada, Marathi, Malayalam, Tamil, and Telegu. It has 10000 speech samples in total having duration of 3-5 seconds recorded in a clean environment. Wavelet transform based scalograms and stft based spectrograms are generated in noisy environment from this speech samples by applying some signal processing techniques which are discussed in detail. These image features are given as input to the CNN model for training. The model is then tested with test samples and the results are compared in terms of recognition accuracies. Later, the performance of the built model is visualized in terms of confusion matrix.