Spoken Language Identification System Using Convolutional Neural Networks

Spoken language identification has gained a lot of importance in the last decade with the advent of text-to-speech systems and intelligent assistants. Language Identification (LID) task revolves around the automated recognition of the language spoken in a given speech utterance. The traditionally used systems in this scenario are GMMs and SVMs. However, these systems also carry disadvantages like limitations in speed and size, discrete data handling, high algorithmic complexity and extensive memory requirements in large scale tasks. Deep Learning, a subset of machine learning, is pushing the boundaries of technological advancements in areas such as spoken language identification. It has already been proven that deep neural networks can be used to efficiently train the speech recognition system. A convolutional neural network refers to a feedforward artificial neural network containing multiple layers of hidden units positioned between its inputs and outputs. In this research work, we have generated melspectrogram images from a collection of speech utterances obtained from the open source website Voxforge. These melspectrograms are used to train the Convolutional Neural Network. The performance of the proposed spoken language identification using three languages namely French, English and Spanish are significantly better than the traditional machine learning approach.